UpdatesV2.json

{
    "internal": [
        [
            "<|BEGIN-VISIBLE-CHAT|>",
            "How can I help you today?"
        ],
        [
            "Hello, how are you functioning today?",
            "I'm functioning well, thank you for asking! I'm here and ready to assist you. How about you? How can I make your day better?"
        ],
        [
            "Okay so in a separate conversation you have helped me write some code for a project.  I've had you write a summary from that conversation and I'll provide it here:\n\n### Summary of Lucid_Autonomy Extension\n\n#### Introduction to Textgen\n\nTextgen is a powerful web-based interface for running large language models (LLMs) locally. It allows users to interact with various LLMs, providing a flexible and customizable environment for text generation tasks. The platform supports a wide range of models and offers numerous extensions to enhance its functionality.\n\n#### Lucid_Autonomy Extension Overview\n\nThe Lucid_Autonomy extension is designed to enable an LLM to interact autonomously with a computer using mouse and keyboard actions. This extension is part of the textgen ecosystem and aims to provide the LLM with \"eyes\" and \"hands\" to perform tasks such as moving the mouse to specific locations on the screen, clicking on elements, entering text, and more.\n\n#### Current Operation of the Extension\n\n1. **Initial Setup and Image Processing**:\n   - The extension takes a screenshot of a specified monitor in a multi-monitor setup.\n   - The screenshot is processed using the Owlv2 model to detect objects based on text queries.\n   - The detected objects are cropped and saved to the `ImageOutputTest` directory.\n   - The cropped images are processed with the MiniCPM-Llama3-V-2_5 model to generate descriptions.\n   - The results are saved in a `results.json` file, containing the image name, coordinates, and description.\n   - A `compressedresults.json` file is created, omitting the coordinates and containing only the image name and description.\n\n2. **LLM Interaction**:\n   - The LLM can interact with the extension using the trigger phrase \"Autonomy_Tasks:\".\n   - If the LLM sends the trigger phrase without any instructions, the extension appends the contents of the `compressedresults.json` file to the LLM's response.\n   - If the LLM sends the trigger phrase with instructions (e.g., `MouseMover = clickable icons_8.png`), the extension executes the specified tasks and then appends the `compressedresults.json` contents to the LLM's response.\n\n3. **Task Execution**:\n   - The extension can handle various tasks such as moving the mouse to specific locations, clicking on elements, entering text, pressing special keys, function keys, key combinations, adding delays, and taking screenshots.\n   - The extension also supports \"Out of Body\" (OOB) commands, which are executed after the LLM's response is complete. These commands are saved to a JSON file and executed with a short delay.\n\n4. **Full Image Processing**:\n   - The extension now processes the entire screenshot with the MiniCPM-Llama3-V-2_5 model using a separate question (`FULL_IMAGE_VISION_MODEL_QUESTION`).\n   - The response from the full image processing is added to the `results.json` and `compressedresults.json` files.\n\n#### Ultimate Goal of the Extension\n\nThe ultimate goal of the Lucid_Autonomy extension is to enable a large language model (LLM) to interact autonomously with a computer. The LLM can perform tasks such as moving the mouse to specific locations on the screen, clicking on elements, entering text, and more. The extension allows the LLM to operate a computer autonomously, effectively giving the LLM \"eyes\" and \"hands\" to interact with the user interface.\n\n\n#### Conclusion\n\nThe Lucid_Autonomy extension is a powerful tool that enables an LLM to interact autonomously with a computer. The current implementation allows the LLM to take screenshots, process images, and move the mouse to specific locations. The ultimate goal is to expand the extension's capabilities to support a wide range of mouse and keyboard interactions, enabling the LLM to perform complex tasks autonomously.\n",
            "Thank you for providing the summary. It's clear and comprehensive. To move forward, how would you like me to assist you with the Lucid_Autonomy extension? Are there specific features you'd like to add, modifications you'd like to make, or perhaps you need help with troubleshooting or optimizing the current implementation? Please let me know how I can best support you."
        ],
        [
            "Let me give you all the code, and then I want you to give me a summary detailing how the code functions and a list of all the current tools at the LLM's disposal:\n\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = \"extensions/Lucid_Autonomy/oob_tasks.json\"\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = \"clickable hyperlinks,clickable icons,text bar field\"  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.2  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = \"Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?\"\nFULL_IMAGE_VISION_MODEL_QUESTION = \"Please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.\"\n\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(\"cuda\")\nelse:\n    device = torch.device(\"cpu\")\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = \"/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5\"\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True)\n    print(\"Owlv2 model loaded.\")\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(\"Owlv2 model unloaded.\")\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if \"int4\" in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 4-bit quantized model loaded on-demand.\")\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 standard model loaded on-demand.\")\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(\"MiniCPM-Llama3 model unloaded.\")\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=\"screenshot.png\"):\n    \"\"\"\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is \"screenshot.png\").\n    \"\"\"\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = \"extensions/Lucid_Autonomy\"\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f\"Screenshot saved as {screenshot_path}\")\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn't exist\n    output_dir = \"extensions/Lucid_Autonomy/ImageOutputTest\"\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f\"Failed to delete {file_path}. Reason: {e}\")\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(\"Processing complete. Check the output directory for the cropped images and results.\")\n\n\n\ndef query_image(img_path, text_queries, score_threshold):\n    \"\"\"\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    \"\"\"\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(\",\")\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=\"pt\").to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][\"boxes\"], results[0][\"scores\"], results[0][\"labels\"]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score < score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        draw.rectangle(box, outline=\"red\", width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=\"red\")\n\n    # Save the annotated image for debugging\n    annotated_image_path = \"extensions/Lucid_Autonomy/annotated_image.png\"\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = \"extensions/Lucid_Autonomy/output.json\"\n    with open(json_output_path, 'w') as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=\"extensions/Lucid_Autonomy/ImageOutputTest\", full_image_response=\"\"):\n    \"\"\"\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is \"extensions/Lucid_Autonomy/ImageOutputTest\").\n        full_image_response (str): The response from the vision model for the full image.\n    \"\"\"\n    with open(json_output_path, 'r') as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f\"{output_dir}/{label}_{i}.png\"\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            \"image_name\": f\"{label}_{i}.png\",\n            \"coordinates\": box,\n            \"description\": vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f\"{output_dir}/results.json\"\n    with open(results_json_path, 'w') as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f\"{output_dir}/compressedresults.json\"\n    compressed_results = [{\"image_name\": result[\"image_name\"], \"description\": result[\"description\"]} for result in results]\n    with open(compressed_results_json_path, 'w') as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            \"image_name\": \"full_image.png\",\n            \"coordinates\": [],\n            \"description\": full_image_response\n        })\n        compressed_results.append({\n            \"image_name\": \"full_image.png\",\n            \"description\": full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, 'w') as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, 'w') as f:\n            json.dump(compressed_results, f, indent=4)\n\n\n\ndef process_with_vision_model(image_path, question):\n    \"\"\"\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    \"\"\"\n    image = Image.open(image_path).convert(\"RGB\")\n    messages = [\n        {\"role\": \"user\", \"content\": f\"{question}\"}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        \"max_new_tokens\": 896,\n        \"repetition_penalty\": 1.05,\n        \"num_beams\": 3,\n        \"top_p\": 0.8,\n        \"top_k\": 1,\n        \"temperature\": 0.7,\n        \"sampling\": True,\n    }\n    if \"int4\" in \"openbmb/MiniCPM-Llama3-V-2_5\":\n        # Disable streaming for the 4-bit model\n        generation_args[\"stream\"] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = \"\"\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end='')\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\n\ndef input_modifier(user_input, state):\n    \"\"\"\n    Modifies the user input before it is processed by the LLM.\n    \"\"\"\n    return user_input\n\ndef output_modifier(output, state):\n    \"\"\"\n    Modifies the LLM output before it is presented in the UI.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the \"Autonomy_Tasks:\" trigger phrase in the LLM's output\n    if \"Autonomy_Tasks:\" in output:\n        # Clean the chat history\n        state[\"history\"] = history_modifier(state[\"history\"])\n        # Extract tasks and execute them\n        tasks_section = output.split(\"Autonomy_Tasks:\")[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n        with open(compressed_results_json_path, 'r') as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, 'w') as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\n\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n\n\n\ndef execute_oob_tasks_with_delay():\n    \"\"\"\n    Executes the saved OOB tasks after a delay.\n    \"\"\"\n    time.sleep(0.1)  # Short delay to ensure the LLM's response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        else:\n            print(f\"Unknown OOB task: {task}\")\n\n\n\n\ndef delay_task(delay):\n    \"\"\"\n    Adds a delay in milliseconds.\n    \"\"\"\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    print(f\"Mouse moved to coordinates: ({x}, {y})\")\n\ndef left_click_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f\"Mouse clicked at coordinates: ({x}, {y})\")\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot.\n    \"\"\"\n    take_screenshot(MONITOR_INDEX, TEXT_QUERIES, SCORE_THRESHOLD, VISION_MODEL_QUESTION, FULL_IMAGE_VISION_MODEL_QUESTION)\n\n\n\ndef move_mouse_to_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef left_click_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    \"\"\"\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor's information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    \"\"\"\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f\"Mouse moved to coordinates: ({center_x}, {center_y})\")\n\ndef get_monitor_info(monitor_index):\n    \"\"\"\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor's information.\n    \"\"\"\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    \"\"\"\n    Modifies the chat history before the text generation in chat mode begins.\n    \"\"\"\n    # Clean the internal history\n    history[\"internal\"] = clean_chat_history(history[\"internal\"])\n    # Clean the visible history\n    history[\"visible\"] = clean_chat_history(history[\"visible\"])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n    with open(compressed_results_json_path, 'r') as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[\"internal\"].append([\"\", f\"<START_COMPRESSED_RESULTS>\\\\\\\\n{latest_compressed_results}\\\\\\\\n<END_COMPRESSED_RESULTS>\"])\n\n    return history\n\ndef clean_chat_history(history):\n    \"\"\"\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) > 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f\"{START_BOOKEND}.*?{END_BOOKEND}\", \"\", history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    \"\"\"\n    Appends the compressed results to the output.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n    output += f\"\\\\\\\\n\\\\\\\\n{START_BOOKEND}\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\n{END_BOOKEND}\"\n    return output\n\ndef text_input(text):\n    \"\"\"\n    Types the given text, replacing '/n' with newline characters.\n    \"\"\"\n    text = text.replace(\"/n\", \"\\\\n\")\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    \"\"\"\n    Presses the specified special key.\n    \"\"\"\n    if key == \"enter\":\n        pyautogui.press(\"enter\")\n    elif key == \"space\":\n        pyautogui.press(\"space\")\n    elif key == \"backspace\":\n        pyautogui.press(\"backspace\")\n    elif key == \"tab\":\n        pyautogui.press(\"tab\")\n    elif key == \"escape\":\n        pyautogui.press(\"esc\")\n    elif key == \"shift\":\n        pyautogui.press(\"shift\")\n    elif key == \"ctrl\":\n        pyautogui.press(\"ctrl\")\n    elif key == \"alt\":\n        pyautogui.press(\"alt\")\n    else:\n        print(f\"Unsupported special key: {key}\")\n\ndef function_key(key):\n    \"\"\"\n    Presses the specified function key.\n    \"\"\"\n    if key.startswith(\"f\") and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f\"Unsupported function key: {key}\")\n\ndef key_combo(combo):\n    \"\"\"\n    Presses the specified key combination.\n    \"\"\"\n    keys = combo.split(\"+\")\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyUp(key)\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()\n",
            "Thank you for providing the code. Here's a summary of how the code functions and a list of the current tools at the LLM's disposal:\n\n### Code Functionality Summary\n\n1. **Imports and Configurations**:\n   - The code imports necessary libraries such as `gradio`, `json`, `pyautogui`, `screeninfo`, `transformers`, `PIL`, `torch`, `os`, `shutil`, `re`, `time`, and `threading`.\n   - Configurable variables include `MONITOR_INDEX`, `TEXT_QUERIES`, `SCORE_THRESHOLD`, `VISION_MODEL_QUESTION`, and `FULL_IMAGE_VISION_MODEL_QUESTION`.\n\n2. **Model Loading and Unloading**:\n   - Functions to load and unload the `Owlv2` model (`load_owlv2_model` and `unload_owlv2_model`).\n   - Functions to load and unload the `MiniCPM-Llama3-V-2_5` model (`load_minicpm_llama_model` and `unload_minicpm_llama_model`).\n\n3. **Screenshot and Image Processing**:\n   - `take_screenshot`: Takes a screenshot of a specified monitor, processes it using the `Owlv2` model to detect objects, and then processes the entire screenshot and cropped images using the `MiniCPM-Llama3-V-2_5` model.\n   - `query_image`: Processes the image using the `Owlv2` model to detect objects based on text queries.\n   - `crop_images`: Crops out the detected objects from the input image based on the bounding box coordinates and processes the cropped images using the `MiniCPM-Llama3-V-2_5` model.\n   - `process_with_vision_model`: Processes the image with the `MiniCPM-Llama3-V-2_5` model and returns the response.\n\n4. **User Interface**:\n   - `ui`: Creates custom Gradio elements for the user interface, allowing users to input parameters and take screenshots.\n\n5. **Input and Output Modifiers**:\n   - `input_modifier`: Modifies the user input before it is processed by the LLM.\n   - `output_modifier`: Modifies the LLM output before it is presented in the UI, handling tasks and appending compressed results.\n\n6. **Task Execution**:\n   - `execute_tasks`: Executes tasks listed in the tasks section, such as moving the mouse, clicking, typing text, pressing special keys, function keys, key combinations, adding delays, and taking screenshots.\n   - `execute_oob_tasks_with_delay` and `execute_oob_tasks`: Handle the execution of out-of-band (OOB) tasks after a delay.\n\n7. **Helper Functions**:\n   - Various helper functions for delaying tasks, moving the mouse, clicking, typing text, pressing special keys, function keys, key combinations, and getting monitor information.\n\n### Current Tools at the LLM's Disposal\n\n1. **Mouse Actions**:\n   - `MouseMover`: Moves the mouse to the center of a specified image.\n   - `MouseClick`: Moves the mouse to specified coordinates or an image and performs a left click.\n   - `MouseMove`: Moves the mouse to specified raw coordinates.\n\n2. **Keyboard Actions**:\n   - `TextInput`: Types the given text, replacing '/n' with newline characters.\n   - `SpecialKey`: Presses specified special keys (e.g., enter, space, backspace, tab, escape, shift, ctrl, alt).\n   - `FunctionKey`: Presses specified function keys (e.g., F1, F2, etc.).\n   - `KeyCombo`: Presses specified key combinations (e.g., ctrl+c, alt+tab).\n\n3. **Delay**:\n   - `Delay`: Adds a delay in milliseconds.\n\n4. **Screenshot**:\n   - `TakeScreenshot`: Takes a screenshot of the specified monitor.\n\n5. **Out-of-Body (OOB) Tasks**:\n   - `OOB_MouseMover`, `OOB_MouseClick`, `OOB_TextInput`, `OOB_SpecialKey`, `OOB_FunctionKey`, `OOB_KeyCombo`, `OOB_Delay`, `OOB_MouseMove`, `OOB_TakeScreenshot`: Similar to the regular tasks but executed after the LLM's response is complete.\n\nThis summary provides an overview of the code's functionality and the tools available to the LLM for autonomous interaction with the computer. If you need further details or assistance with specific parts of the code, feel free to ask!"
        ],
        [
            "Okay so you can see that there are gradio elements with text fields I can changes.  These elements work when I as the user press the \"Take Screenshot\" button, but when the LLM sends a command like this:\n\n```\nAutonomy_Tasks:\nTakeScreenshot\\\n```\n\nOnly the hard coded values from the extension code are used and not the information from the text fields when I update them.",
            "I understand the issue. The current implementation of the `take_screenshot` function uses hardcoded values for `MONITOR_INDEX`, `TEXT_QUERIES`, `SCORE_THRESHOLD`, `VISION_MODEL_QUESTION`, and `FULL_IMAGE_VISION_MODEL_QUESTION`. When the LLM sends a command like `TakeScreenshot`, it doesn't take into account the updated values from the Gradio text fields.\n\nTo fix this, we need to modify the `take_screenshot_task` function to use the current values from the Gradio interface instead of the hardcoded values. Here's how you can do it:\n\n1. **Store the Gradio interface inputs in global variables**: Update the global variables whenever the user changes the input fields in the Gradio interface.\n\n2. **Modify the `take_screenshot_task` function**: Make the `take_screenshot_task` function use these global variables instead of the hardcoded values.\n\nHere's the updated code:\n\n```python\n# Global variables for Gradio inputs\nglobal_monitor_index = MONITOR_INDEX\nglobal_text_queries = TEXT_QUERIES\nglobal_score_threshold = SCORE_THRESHOLD\nglobal_vision_model_question = VISION_MODEL_QUESTION\nglobal_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: setattr(global_vars, 'global_monitor_index', int(x)), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: setattr(global_vars, 'global_text_queries', x), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: setattr(global_vars, 'global_score_threshold', float(x)), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: setattr(global_vars, 'global_vision_model_question', x), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: setattr(global_vars, 'global_full_image_vision_model_question', x), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot using the current values from the Gradio interface.\n    \"\"\"\n    take_screenshot(global_monitor_index, global_text_queries, global_score_threshold, global_vision_model_question, global_full_image_vision_model_question)\n\n# Define a class to hold global variables\nclass GlobalVars:\n    pass\n\nglobal_vars = GlobalVars()\nglobal_vars.global_monitor_index = MONITOR_INDEX\nglobal_vars.global_text_queries = TEXT_QUERIES\nglobal_vars.global_score_threshold = SCORE_THRESHOLD\nglobal_vars.global_vision_model_question = VISION_MODEL_QUESTION\nglobal_vars.global_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()\n```\n\nIn this updated code:\n- We define global variables to store the current values from the Gradio interface.\n- We update these global variables whenever the user changes the input fields in the Gradio interface.\n- The `take_screenshot_task` function now uses these global variables instead of the hardcoded values.\n\nThis should ensure that the `TakeScreenshot` command uses the updated values from the Gradio text fields."
        ],
        [
            "Unfortunately that did not work.  Can you look at the code in its current state, maybe I implemented your change incorrectly?  As a user using the UI I can press the \"Take Screenshot\" button and the changes I make to things like \"Score Threshold\" are overtly working (fewer icons are being picked up by owl).  But when I then ask the LLM to take a screenshot, it sends the right commands, a screenshot is taken, but I can tell the values in the UI that I changed are not making their way to the LLM's utilization of \"Take Screenshot\" because there are a lot of icons recognized with the default value of 0.2.  So I know the LLM Autonomy_Task implementation of \"Take Screenshot\" is using the default value of 0.2 and not the changed value that I put in the UI.\n\nCurrent state of code:\n\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = \"extensions/Lucid_Autonomy/oob_tasks.json\"\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = \"clickable hyperlinks,clickable icons,text bar field\"  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.2  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = \"Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?\"\nFULL_IMAGE_VISION_MODEL_QUESTION = \"Please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.\"\n\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(\"cuda\")\nelse:\n    device = torch.device(\"cpu\")\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = \"/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5\"\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True)\n    print(\"Owlv2 model loaded.\")\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(\"Owlv2 model unloaded.\")\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if \"int4\" in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 4-bit quantized model loaded on-demand.\")\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 standard model loaded on-demand.\")\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(\"MiniCPM-Llama3 model unloaded.\")\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=\"screenshot.png\"):\n    \"\"\"\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is \"screenshot.png\").\n    \"\"\"\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = \"extensions/Lucid_Autonomy\"\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f\"Screenshot saved as {screenshot_path}\")\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn't exist\n    output_dir = \"extensions/Lucid_Autonomy/ImageOutputTest\"\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f\"Failed to delete {file_path}. Reason: {e}\")\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(\"Processing complete. Check the output directory for the cropped images and results.\")\n\n\n\ndef query_image(img_path, text_queries, score_threshold):\n    \"\"\"\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    \"\"\"\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(\",\")\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=\"pt\").to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][\"boxes\"], results[0][\"scores\"], results[0][\"labels\"]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score < score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        draw.rectangle(box, outline=\"red\", width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=\"red\")\n\n    # Save the annotated image for debugging\n    annotated_image_path = \"extensions/Lucid_Autonomy/annotated_image.png\"\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = \"extensions/Lucid_Autonomy/output.json\"\n    with open(json_output_path, 'w') as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=\"extensions/Lucid_Autonomy/ImageOutputTest\", full_image_response=\"\"):\n    \"\"\"\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is \"extensions/Lucid_Autonomy/ImageOutputTest\").\n        full_image_response (str): The response from the vision model for the full image.\n    \"\"\"\n    with open(json_output_path, 'r') as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f\"{output_dir}/{label}_{i}.png\"\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            \"image_name\": f\"{label}_{i}.png\",\n            \"coordinates\": box,\n            \"description\": vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f\"{output_dir}/results.json\"\n    with open(results_json_path, 'w') as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f\"{output_dir}/compressedresults.json\"\n    compressed_results = [{\"image_name\": result[\"image_name\"], \"description\": result[\"description\"]} for result in results]\n    with open(compressed_results_json_path, 'w') as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            \"image_name\": \"full_image.png\",\n            \"coordinates\": [],\n            \"description\": full_image_response\n        })\n        compressed_results.append({\n            \"image_name\": \"full_image.png\",\n            \"description\": full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, 'w') as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, 'w') as f:\n            json.dump(compressed_results, f, indent=4)\n\n\n\ndef process_with_vision_model(image_path, question):\n    \"\"\"\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    \"\"\"\n    image = Image.open(image_path).convert(\"RGB\")\n    messages = [\n        {\"role\": \"user\", \"content\": f\"{question}\"}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        \"max_new_tokens\": 896,\n        \"repetition_penalty\": 1.05,\n        \"num_beams\": 3,\n        \"top_p\": 0.8,\n        \"top_k\": 1,\n        \"temperature\": 0.7,\n        \"sampling\": True,\n    }\n    if \"int4\" in \"openbmb/MiniCPM-Llama3-V-2_5\":\n        # Disable streaming for the 4-bit model\n        generation_args[\"stream\"] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = \"\"\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end='')\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_monitor_index = MONITOR_INDEX\nglobal_text_queries = TEXT_QUERIES\nglobal_score_threshold = SCORE_THRESHOLD\nglobal_vision_model_question = VISION_MODEL_QUESTION\nglobal_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: setattr(global_vars, 'global_monitor_index', int(x)), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: setattr(global_vars, 'global_text_queries', x), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: setattr(global_vars, 'global_score_threshold', float(x)), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: setattr(global_vars, 'global_vision_model_question', x), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: setattr(global_vars, 'global_full_image_vision_model_question', x), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\n\ndef input_modifier(user_input, state):\n    \"\"\"\n    Modifies the user input before it is processed by the LLM.\n    \"\"\"\n    return user_input\n\ndef output_modifier(output, state):\n    \"\"\"\n    Modifies the LLM output before it is presented in the UI.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the \"Autonomy_Tasks:\" trigger phrase in the LLM's output\n    if \"Autonomy_Tasks:\" in output:\n        # Clean the chat history\n        state[\"history\"] = history_modifier(state[\"history\"])\n        # Extract tasks and execute them\n        tasks_section = output.split(\"Autonomy_Tasks:\")[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n        with open(compressed_results_json_path, 'r') as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, 'w') as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\n\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n\n\n\ndef execute_oob_tasks_with_delay():\n    \"\"\"\n    Executes the saved OOB tasks after a delay.\n    \"\"\"\n    time.sleep(0.1)  # Short delay to ensure the LLM's response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        else:\n            print(f\"Unknown OOB task: {task}\")\n\n\n\n\ndef delay_task(delay):\n    \"\"\"\n    Adds a delay in milliseconds.\n    \"\"\"\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    print(f\"Mouse moved to coordinates: ({x}, {y})\")\n\ndef left_click_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f\"Mouse clicked at coordinates: ({x}, {y})\")\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot using the current values from the Gradio interface.\n    \"\"\"\n    take_screenshot(global_monitor_index, global_text_queries, global_score_threshold, global_vision_model_question, global_full_image_vision_model_question)\n\n# Define a class to hold global variables\nclass GlobalVars:\n    pass\n\nglobal_vars = GlobalVars()\nglobal_vars.global_monitor_index = MONITOR_INDEX\nglobal_vars.global_text_queries = TEXT_QUERIES\nglobal_vars.global_score_threshold = SCORE_THRESHOLD\nglobal_vars.global_vision_model_question = VISION_MODEL_QUESTION\nglobal_vars.global_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\n\ndef move_mouse_to_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef left_click_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    \"\"\"\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor's information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    \"\"\"\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f\"Mouse moved to coordinates: ({center_x}, {center_y})\")\n\ndef get_monitor_info(monitor_index):\n    \"\"\"\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor's information.\n    \"\"\"\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    \"\"\"\n    Modifies the chat history before the text generation in chat mode begins.\n    \"\"\"\n    # Clean the internal history\n    history[\"internal\"] = clean_chat_history(history[\"internal\"])\n    # Clean the visible history\n    history[\"visible\"] = clean_chat_history(history[\"visible\"])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n    with open(compressed_results_json_path, 'r') as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[\"internal\"].append([\"\", f\"<START_COMPRESSED_RESULTS>\\\\\\\\n{latest_compressed_results}\\\\\\\\n<END_COMPRESSED_RESULTS>\"])\n\n    return history\n\ndef clean_chat_history(history):\n    \"\"\"\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) > 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f\"{START_BOOKEND}.*?{END_BOOKEND}\", \"\", history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    \"\"\"\n    Appends the compressed results to the output.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n    output += f\"\\\\\\\\n\\\\\\\\n{START_BOOKEND}\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\n{END_BOOKEND}\"\n    return output\n\ndef text_input(text):\n    \"\"\"\n    Types the given text, replacing '/n' with newline characters.\n    \"\"\"\n    text = text.replace(\"/n\", \"\\\\n\")\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    \"\"\"\n    Presses the specified special key.\n    \"\"\"\n    if key == \"enter\":\n        pyautogui.press(\"enter\")\n    elif key == \"space\":\n        pyautogui.press(\"space\")\n    elif key == \"backspace\":\n        pyautogui.press(\"backspace\")\n    elif key == \"tab\":\n        pyautogui.press(\"tab\")\n    elif key == \"escape\":\n        pyautogui.press(\"esc\")\n    elif key == \"shift\":\n        pyautogui.press(\"shift\")\n    elif key == \"ctrl\":\n        pyautogui.press(\"ctrl\")\n    elif key == \"alt\":\n        pyautogui.press(\"alt\")\n    else:\n        print(f\"Unsupported special key: {key}\")\n\ndef function_key(key):\n    \"\"\"\n    Presses the specified function key.\n    \"\"\"\n    if key.startswith(\"f\") and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f\"Unsupported function key: {key}\")\n\ndef key_combo(combo):\n    \"\"\"\n    Presses the specified key combination.\n    \"\"\"\n    keys = combo.split(\"+\")\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyUp(key)\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()\n",
            "Thank you for providing the current state of your code. I see the issue now. The global variables are not being updated correctly because the `setattr` function is not being used properly. Additionally, the `GlobalVars` class is not being utilized correctly.\n\nLet's correct the implementation to ensure that the global variables are updated properly when the user changes the input fields in the Gradio interface. Here's the updated code:\n\n```python\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = \"extensions/Lucid_Autonomy/oob_tasks.json\"\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = \"clickable hyperlinks,clickable icons,text bar field\"  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.2  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = \"Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?\"\nFULL_IMAGE_VISION_MODEL_QUESTION = \"Please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.\"\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(\"cuda\")\nelse:\n    device = torch.device(\"cpu\")\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = \"/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5\"\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True)\n    print(\"Owlv2 model loaded.\")\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(\"Owlv2 model unloaded.\")\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if \"int4\" in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 4-bit quantized model loaded on-demand.\")\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 standard model loaded on-demand.\")\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(\"MiniCPM-Llama3 model unloaded.\")\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=\"screenshot.png\"):\n    \"\"\"\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is \"screenshot.png\").\n    \"\"\"\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = \"extensions/Lucid_Autonomy\"\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f\"Screenshot saved as {screenshot_path}\")\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn't exist\n    output_dir = \"extensions/Lucid_Autonomy/ImageOutputTest\"\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f\"Failed to delete {file_path}. Reason: {e}\")\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(\"Processing complete. Check the output directory for the cropped images and results.\")\n\ndef query_image(img_path, text_queries, score_threshold):\n    \"\"\"\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    \"\"\"\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(\",\")\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=\"pt\").to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][\"boxes\"], results[0][\"scores\"], results[0][\"labels\"]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score < score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        draw.rectangle(box, outline=\"red\", width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=\"red\")\n\n    # Save the annotated image for debugging\n    annotated_image_path = \"extensions/Lucid_Autonomy/annotated_image.png\"\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = \"extensions/Lucid_Autonomy/output.json\"\n    with open(json_output_path, 'w') as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=\"extensions/Lucid_Autonomy/ImageOutputTest\", full_image_response=\"\"):\n    \"\"\"\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is \"extensions/Lucid_Autonomy/ImageOutputTest\").\n        full_image_response (str): The response from the vision model for the full image.\n    \"\"\"\n    with open(json_output_path, 'r') as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f\"{output_dir}/{label}_{i}.png\"\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            \"image_name\": f\"{label}_{i}.png\",\n            \"coordinates\": box,\n            \"description\": vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f\"{output_dir}/results.json\"\n    with open(results_json_path, 'w') as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f\"{output_dir}/compressedresults.json\"\n    compressed_results = [{\"image_name\": result[\"image_name\"], \"description\": result[\"description\"]} for result in results]\n    with open(compressed_results_json_path, 'w') as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            \"image_name\": \"full_image.png\",\n            \"coordinates\": [],\n            \"description\": full_image_response\n        })\n        compressed_results.append({\n            \"image_name\": \"full_image.png\",\n            \"description\": full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, 'w') as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, 'w') as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    \"\"\"\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    \"\"\"\n    image = Image.open(image_path).convert(\"RGB\")\n    messages = [\n        {\"role\": \"user\", \"content\": f\"{question}\"}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        \"max_new_tokens\": 896,\n        \"repetition_penalty\": 1.05,\n        \"num_beams\": 3,\n        \"top_p\": 0.8,\n        \"top_k\": 1,\n        \"temperature\": 0.7,\n        \"sampling\": True,\n    }\n    if \"int4\" in \"openbmb/MiniCPM-Llama3-V-2_5\":\n        # Disable streaming for the 4-bit model\n        generation_args[\"stream\"] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = \"\"\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end='')\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    \"global_monitor_index\": MONITOR_INDEX,\n    \"global_text_queries\": TEXT_QUERIES,\n    \"global_score_threshold\": SCORE_THRESHOLD,\n    \"global_vision_model_question\": VISION_MODEL_QUESTION,\n    \"global_full_image_vision_model_question\": FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({\"global_monitor_index\": int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({\"global_text_queries\": x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({\"global_score_threshold\": float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({\"global_vision_model_question\": x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({\"global_full_image_vision_model_question\": x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    \"\"\"\n    Modifies the user input before it is processed by the LLM.\n    \"\"\"\n    return user_input\n\ndef output_modifier(output, state):\n    \"\"\"\n    Modifies the LLM output before it is presented in the UI.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the \"Autonomy_Tasks:\" trigger phrase in the LLM's output\n    if \"Autonomy_Tasks:\" in output:\n        # Clean the chat history\n        state[\"history\"] = history_modifier(state[\"history\"])\n        # Extract tasks and execute them\n        tasks_section = output.split(\"Autonomy_Tasks:\")[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n        with open(compressed_results_json_path, 'r') as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, 'w') as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n\ndef execute_oob_tasks_with_delay():\n    \"\"\"\n    Executes the saved OOB tasks after a delay.\n    \"\"\"\n    time.sleep(0.1)  # Short delay to ensure the LLM's response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        else:\n            print(f\"Unknown OOB task: {task}\")\n\ndef delay_task(delay):\n    \"\"\"\n    Adds a delay in milliseconds.\n    \"\"\"\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    print(f\"Mouse moved to coordinates: ({x}, {y})\")\n\ndef left_click_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f\"Mouse clicked at coordinates: ({x}, {y})\")\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot using the current values from the Gradio interface.\n    \"\"\"\n    take_screenshot(global_vars[\"global_monitor_index\"], global_vars[\"global_text_queries\"], global_vars[\"global_score_threshold\"], global_vars[\"global_vision_model_question\"], global_vars[\"global_full_image_vision_model_question\"])\n\ndef move_mouse_to_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef left_click_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    \"\"\"\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor's information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    \"\"\"\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f\"Mouse moved to coordinates: ({center_x}, {center_y})\")\n\ndef get_monitor_info(monitor_index):\n    \"\"\"\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor's information.\n    \"\"\"\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    \"\"\"\n    Modifies the chat history before the text generation in chat mode begins.\n    \"\"\"\n    # Clean the internal history\n    history[\"internal\"] = clean_chat_history(history[\"internal\"])\n    # Clean the visible history\n    history[\"visible\"] = clean_chat_history(history[\"visible\"])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n    with open(compressed_results_json_path, 'r') as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[\"internal\"].append([\"\", f\"<START_COMPRESSED_RESULTS>\\\\\\\\n{latest_compressed_results}\\\\\\\\n<END_COMPRESSED_RESULTS>\"])\n\n    return history\n\ndef clean_chat_history(history):\n    \"\"\"\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) > 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f\"{START_BOOKEND}.*?{END_BOOKEND}\", \"\", history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    \"\"\"\n    Appends the compressed results to the output.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n    output += f\"\\\\\\\\n\\\\\\\\n{START_BOOKEND}\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\n{END_BOOKEND}\"\n    return output\n\ndef text_input(text):\n    \"\"\"\n    Types the given text, replacing '/n' with newline characters.\n    \"\"\"\n    text = text.replace(\"/n\", \"\\\\n\")\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    \"\"\"\n    Presses the specified special key.\n    \"\"\"\n    if key == \"enter\":\n        pyautogui.press(\"enter\")\n    elif key == \"space\":\n        pyautogui.press(\"space\")\n    elif key == \"backspace\":\n        pyautogui.press(\"backspace\")\n    elif key == \"tab\":\n        pyautogui.press(\"tab\")\n    elif key == \"escape\":\n        pyautogui.press(\"esc\")\n    elif key == \"shift\":\n        pyautogui.press(\"shift\")\n    elif key == \"ctrl\":\n        pyautogui.press(\"ctrl\")\n    elif key == \"alt\":\n        pyautogui.press(\"alt\")\n    else:\n        print(f\"Unsupported special key: {key}\")\n\ndef function_key(key):\n    \"\"\"\n    Presses the specified function key.\n    \"\"\"\n    if key.startswith(\"f\") and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f\"Unsupported function key: {key}\")\n\ndef key_combo(combo):\n    \"\"\"\n    Presses the specified key combination.\n    \"\"\"\n    keys = combo.split(\"+\")\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyUp(key)\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()\n```\n\nIn this updated code, I've corrected the implementation to ensure that the global variables are updated properly when the user changes the input fields in the Gradio interface. The `global_vars` dictionary is used to store the current values from the Gradio interface, and the `take_screenshot_task` function now uses these global variables instead of the hardcoded values. This should ensure that the `TakeScreenshot` command uses the updated values from the Gradio text fields."
        ],
        [
            "Yeass! Good job :3 *head pats*  okay let's add one more function please then hopefully I can call it a night, I'm tired but this is so much fun!\n\nOkay this one is hopefully simple, can you add a tool that lets the LLM page up and page down? Don't write any code, can you also explain the key combo section to me what does keyDown and keyUp mean in this context?",
            "I'm glad to hear that you're enjoying this! Let's add the functionality for the LLM to page up and page down.\n\n### Explanation of `keyDown` and `keyUp`\n\nIn the context of the `key_combo` function, `keyDown` and `keyUp` are used to simulate pressing and releasing keys on the keyboard. This is particularly useful for key combinations where you need to hold down one or more keys while pressing another key.\n\n- **`keyDown(key)`**: Simulates pressing down a key and holding it.\n- **`keyUp(key)`**: Simulates releasing a key that was previously pressed down.\n\nFor example, if you want to simulate pressing `Ctrl + C` (copy), you would:\n1. Press down the `Ctrl` key using `keyDown(\"ctrl\")`.\n2. Press the `C` key using `pyautogui.press(\"c\")`.\n3. Release the `Ctrl` key using `keyUp(\"ctrl\")`.\n\nThis sequence ensures that the `Ctrl` key is held down while the `C` key is pressed, effectively performing the copy operation.\n\n### Adding Page Up and Page Down Functionality\n\nTo add the functionality for the LLM to page up and page down, you can extend the `execute_tasks` function to handle new tasks `PageUp` and `PageDown`. These tasks will use the `pyautogui.press` function to simulate pressing the `Page Up` and `Page Down` keys.\n\nHere's how you can do it:\n\n1. **Extend the `execute_tasks` function**: Add new cases for `PageUp` and `PageDown`.\n2. **Update the `key_combo` function**: Ensure it can handle the new key combinations if needed.\n\n### Updated `execute_tasks` Function\n\n```python\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name"
        ],
        [
            "Sorry I had to stop you mid reply, but do we need to also update the execute_oob_tasks?  If so can you really review the working code closely to make sure you caught and contextualized everything that needs changing?",
            "Absolutely, let's ensure that we update both the `execute_tasks` and `execute_oob_tasks` functions to handle the new `PageUp` and `PageDown` tasks. I'll also review the working code closely to make sure everything is contextualized correctly.\n\n### Updated `execute_tasks` Function\n\n```python\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"PageDown\":\n            pyautogui.press(\"pagedown\")\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n```\n\n### Updated `execute_oob_tasks` Function\n\n```python\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"OOB_PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"OOB_PageDown\":\n            pyautogui.press(\"pagedown\")\n        else:\n            print(f\"Unknown OOB task: {task}\")\n```\n\n### Explanation of `keyDown` and `keyUp`\n\nIn the context of the `key_combo` function, `keyDown` and `keyUp` are used to simulate pressing and releasing keys on the keyboard. This is particularly useful for key combinations where you need to hold down one or more keys while pressing another key.\n\n- **`keyDown(key)`**: Simulates pressing down a key and holding it.\n- **`keyUp(key)`**: Simulates releasing a key that was previously pressed down.\n\nFor example, if you want to simulate pressing `Ctrl + C` (copy), you would:\n1. Press down the `Ctrl` key using `keyDown(\"ctrl\")`.\n2. Press the `C` key using `pyautogui.press(\"c\")`.\n3. Release the `Ctrl` key using `keyUp(\"ctrl\")`.\n\nThis sequence ensures that the `Ctrl` key is held down while the `C` key is pressed, effectively performing the copy operation.\n\n### Summary\n\nWith these updates, the LLM can now handle `PageUp` and `PageDown` tasks, both in the immediate execution and in the out-of-band (OOB) tasks. This should provide the functionality you need for paging up and down. If you have any further questions or need additional modifications, feel free to ask!"
        ],
        [
            "Okay, the code has been working perfectly for the last couple of days, I've been testing it out with a different model and it is surfing the internet and everything! :3 omg it's amazing!  There are a few things I would like to change at this point however.\n\nOkay when the boxes are drawn on the screenshot and the annotated_image.png is created, sometimes there are too many red lines forming the boxes, which obscures the cropped out images for the minicpm vision model.\n\nSo as you'll see in the code below, I've commented out the line that draws the boxes.  However, me being able to see the boxes is very useful.  So I would like you to keep the commented out red boxes (maybe I'll turn that back on at some point) and keep the annoated_image.png going to the minicpm model as we have it, but I want you to create a new image called Human_annotated_image.png, I want it created in the same directory as the annotated_image.png.  This way I can look at the Human_annotated_image.png and the Vision AI can use the annotated_image.png\n\nI have an additional change after this, but it is more complex and we'll just work on one at a time.\n\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = \"extensions/Lucid_Autonomy/oob_tasks.json\"\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = \"clickable text hyperlinks,clickable icons,text bar field,clickable UI buttons,UI tabs\"  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.20  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = \"Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?\"\nFULL_IMAGE_VISION_MODEL_QUESTION = \"This is a computer screenshot, please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.  You are aiding in helping those with visual handicaps.\"\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(\"cuda\")\nelse:\n    device = torch.device(\"cpu\")\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = \"/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5\"\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True)\n    print(\"Owlv2 model loaded.\")\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(\"Owlv2 model unloaded.\")\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if \"int4\" in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 4-bit quantized model loaded on-demand.\")\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 standard model loaded on-demand.\")\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(\"MiniCPM-Llama3 model unloaded.\")\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=\"screenshot.png\"):\n    \"\"\"\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is \"screenshot.png\").\n    \"\"\"\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = \"extensions/Lucid_Autonomy\"\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f\"Screenshot saved as {screenshot_path}\")\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn't exist\n    output_dir = \"extensions/Lucid_Autonomy/ImageOutputTest\"\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f\"Failed to delete {file_path}. Reason: {e}\")\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(\"Processing complete. Check the output directory for the cropped images and results.\")\n\ndef query_image(img_path, text_queries, score_threshold):\n    \"\"\"\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    \"\"\"\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(\",\")\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=\"pt\").to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][\"boxes\"], results[0][\"scores\"], results[0][\"labels\"]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score < score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        #draw.rectangle(box, outline=\"red\", width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=\"red\")\n\n    # Save the annotated image for debugging\n    annotated_image_path = \"extensions/Lucid_Autonomy/annotated_image.png\"\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = \"extensions/Lucid_Autonomy/output.json\"\n    with open(json_output_path, 'w') as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=\"extensions/Lucid_Autonomy/ImageOutputTest\", full_image_response=\"\"):\n    \"\"\"\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is \"extensions/Lucid_Autonomy/ImageOutputTest\").\n        full_image_response (str): The response from the vision model for the full image.\n    \"\"\"\n    with open(json_output_path, 'r') as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f\"{output_dir}/{label}_{i}.png\"\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            \"image_name\": f\"{label}_{i}.png\",\n            \"coordinates\": box,\n            \"description\": vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f\"{output_dir}/results.json\"\n    with open(results_json_path, 'w') as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f\"{output_dir}/compressedresults.json\"\n    compressed_results = [{\"image_name\": result[\"image_name\"], \"description\": result[\"description\"]} for result in results]\n    with open(compressed_results_json_path, 'w') as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            \"image_name\": \"full_image.png\",\n            \"coordinates\": [],\n            \"description\": full_image_response\n        })\n        compressed_results.append({\n            \"image_name\": \"full_image.png\",\n            \"description\": full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, 'w') as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, 'w') as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    \"\"\"\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    \"\"\"\n    image = Image.open(image_path).convert(\"RGB\")\n    messages = [\n        {\"role\": \"user\", \"content\": f\"{question}\"}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        \"max_new_tokens\": 896,\n        \"repetition_penalty\": 1.05,\n        \"num_beams\": 3,\n        \"top_p\": 0.8,\n        \"top_k\": 1,\n        \"temperature\": 0.7,\n        \"sampling\": True,\n    }\n    if \"int4\" in \"openbmb/MiniCPM-Llama3-V-2_5\":\n        # Disable streaming for the 4-bit model\n        generation_args[\"stream\"] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = \"\"\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end='')\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    \"global_monitor_index\": MONITOR_INDEX,\n    \"global_text_queries\": TEXT_QUERIES,\n    \"global_score_threshold\": SCORE_THRESHOLD,\n    \"global_vision_model_question\": VISION_MODEL_QUESTION,\n    \"global_full_image_vision_model_question\": FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({\"global_monitor_index\": int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({\"global_text_queries\": x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({\"global_score_threshold\": float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({\"global_vision_model_question\": x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({\"global_full_image_vision_model_question\": x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    \"\"\"\n    Modifies the user input before it is processed by the LLM.\n    \"\"\"\n    return user_input\n\ndef output_modifier(output, state):\n    \"\"\"\n    Modifies the LLM output before it is presented in the UI.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the \"Autonomy_Tasks:\" trigger phrase in the LLM's output\n    if \"Autonomy_Tasks:\" in output:\n        # Clean the chat history\n        state[\"history\"] = history_modifier(state[\"history\"])\n        # Extract tasks and execute them\n        tasks_section = output.split(\"Autonomy_Tasks:\")[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n        with open(compressed_results_json_path, 'r') as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, 'w') as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"PageDown\":\n            pyautogui.press(\"pagedown\")\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n\n\ndef execute_oob_tasks_with_delay():\n    \"\"\"\n    Executes the saved OOB tasks after a delay.\n    \"\"\"\n    time.sleep(0.1)  # Short delay to ensure the LLM's response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"OOB_PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"OOB_PageDown\":\n            pyautogui.press(\"pagedown\")\n        else:\n            print(f\"Unknown OOB task: {task}\")\n\n\ndef delay_task(delay):\n    \"\"\"\n    Adds a delay in milliseconds.\n    \"\"\"\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    print(f\"Mouse moved to coordinates: ({x}, {y})\")\n\ndef left_click_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f\"Mouse clicked at coordinates: ({x}, {y})\")\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot using the current values from the Gradio interface.\n    \"\"\"\n    take_screenshot(global_vars[\"global_monitor_index\"], global_vars[\"global_text_queries\"], global_vars[\"global_score_threshold\"], global_vars[\"global_vision_model_question\"], global_vars[\"global_full_image_vision_model_question\"])\n\ndef move_mouse_to_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef left_click_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    \"\"\"\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor's information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    \"\"\"\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f\"Mouse moved to coordinates: ({center_x}, {center_y})\")\n\ndef get_monitor_info(monitor_index):\n    \"\"\"\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor's information.\n    \"\"\"\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    \"\"\"\n    Modifies the chat history before the text generation in chat mode begins.\n    \"\"\"\n    # Clean the internal history\n    history[\"internal\"] = clean_chat_history(history[\"internal\"])\n    # Clean the visible history\n    history[\"visible\"] = clean_chat_history(history[\"visible\"])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n    with open(compressed_results_json_path, 'r') as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[\"internal\"].append([\"\", f\"<START_COMPRESSED_RESULTS>\\\\\\\\\\\\\\\\n{latest_compressed_results}\\\\\\\\\\\\\\\\n<END_COMPRESSED_RESULTS>\"])\n\n    return history\n\ndef clean_chat_history(history):\n    \"\"\"\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) > 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f\"{START_BOOKEND}.*?{END_BOOKEND}\", \"\", history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    \"\"\"\n    Appends the compressed results to the output.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n    output += f\"\\\\\\\\\\\\\\\\n\\\\\\\\\\\\\\\\n{START_BOOKEND}\\\\\\\\\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\\\\\\\\\n{END_BOOKEND}\"\n    return output\n\ndef text_input(text):\n    \"\"\"\n    Types the given text, replacing '/n' with newline characters.\n    \"\"\"\n    text = text.replace(\"/n\", \"\\\\\\\\n\")\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    \"\"\"\n    Presses the specified special key.\n    \"\"\"\n    if key == \"enter\":\n        pyautogui.press(\"enter\")\n    elif key == \"space\":\n        pyautogui.press(\"space\")\n    elif key == \"backspace\":\n        pyautogui.press(\"backspace\")\n    elif key == \"tab\":\n        pyautogui.press(\"tab\")\n    elif key == \"escape\":\n        pyautogui.press(\"esc\")\n    elif key == \"shift\":\n        pyautogui.press(\"shift\")\n    elif key == \"ctrl\":\n        pyautogui.press(\"ctrl\")\n    elif key == \"alt\":\n        pyautogui.press(\"alt\")\n    else:\n        print(f\"Unsupported special key: {key}\")\n\ndef function_key(key):\n    \"\"\"\n    Presses the specified function key.\n    \"\"\"\n    if key.startswith(\"f\") and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f\"Unsupported function key: {key}\")\n\ndef key_combo(combo):\n    \"\"\"\n    Presses the specified key combination.\n    \"\"\"\n    keys = combo.split(\"+\")\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyUp(key)\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()",
            "Sure, let's add the functionality to create a new image called `Human_annotated_image.png` that includes the red boxes for human inspection, while keeping the `annotated_image.png` without the red boxes for the vision AI model.\n\nHere's the updated code with the requested changes:\n\n```python\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = \"extensions/Lucid_Autonomy/oob_tasks.json\"\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = \"clickable text hyperlinks,clickable icons,text bar field,clickable UI buttons,UI tabs\"  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.20  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = \"Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?\"\nFULL_IMAGE_VISION_MODEL_QUESTION = \"This is a computer screenshot, please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.  You are aiding in helping those with visual handicaps.\"\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(\"cuda\")\nelse:\n    device = torch.device(\"cpu\")\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = \"/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5\"\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True)\n    print(\"Owlv2 model loaded.\")\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(\"Owlv2 model unloaded.\")\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if \"int4\" in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 4-bit quantized model loaded on-demand.\")\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 standard model loaded on-demand.\")\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(\"MiniCPM-Llama3 model unloaded.\")\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=\"screenshot.png\"):\n    \"\"\"\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is \"screenshot.png\").\n    \"\"\"\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = \"extensions/Lucid_Autonomy\"\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f\"Screenshot saved as {screenshot_path}\")\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, human_annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(annotated_image_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn't exist\n    output_dir = \"extensions/Lucid_Autonomy/ImageOutputTest\"\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f\"Failed to delete {file_path}. Reason: {e}\")\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(\"Processing complete. Check the output directory for the cropped images and results.\")\n\ndef query_image(img_path, text_queries, score_threshold):\n    \"\"\"\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    \"\"\"\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(\",\")\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=\"pt\").to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][\"boxes\"], results[0][\"scores\"], results[0][\"labels\"]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    human_img_pil = img.copy()\n    human_draw = ImageDraw.Draw(human_img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score < score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=\"red\")\n\n        # Draw the boxes for the human-annotated image\n        human_draw.rectangle(box, outline=\"red\", width=2)\n\n    # Save the annotated image for debugging\n    annotated_image_path = \"extensions/Lucid_Autonomy/annotated_image.png\"\n    img_pil.save(annotated_image_path)\n\n    # Save the human-annotated image for human inspection\n    human_annotated_image_path = \"extensions/Lucid_Autonomy/Human_annotated_image.png\"\n    human_img_pil.save(human_annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = \"extensions/Lucid_Autonomy/output.json\"\n    with open(json_output_path, 'w') as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, human_annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=\"extensions/Lucid_Autonomy/ImageOutputTest\", full_image_response=\"\"):\n    \"\"\"\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is \"extensions/Lucid_Autonomy/ImageOutputTest\").\n        full_image_response (str): The response from the vision model for the full image.\n    \"\"\"\n    with open(json_output_path, 'r') as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f\"{output_dir}/{label}_{i}.png\"\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            \"image_name\": f\"{label}_{i}.png\",\n            \"coordinates\": box,\n            \"description\": vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f\"{output_dir}/results.json\"\n    with open(results_json_path, 'w') as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f\"{output_dir}/compressedresults.json\"\n    compressed_results = [{\"image_name\": result[\"image_name\"], \"description\": result[\"description\"]} for result in results]\n    with open(compressed_results_json_path, 'w') as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            \"image_name\": \"full_image.png\",\n            \"coordinates\": [],\n            \"description\": full_image_response\n        })\n        compressed_results.append({\n            \"image_name\": \"full_image.png\",\n            \"description\": full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, 'w') as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, 'w') as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    \"\"\"\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    \"\"\"\n    image = Image.open(image_path).convert(\"RGB\")\n    messages = [\n        {\"role\": \"user\", \"content\": f\"{question}\"}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        \"max_new_tokens\": 896,\n        \"repetition_penalty\": 1.05,\n        \"num_beams\": 3,\n        \"top_p\": 0.8,\n        \"top_k\": 1,\n        \"temperature\": 0.7,\n        \"sampling\": True,\n    }\n    if \"int4\" in \"openbmb/MiniCPM-Llama3-V-2_5\":\n        # Disable streaming for the 4-bit model\n        generation_args[\"stream\"] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = \"\"\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end='')\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    \"global_monitor_index\": MONITOR_INDEX,\n    \"global_text_queries\": TEXT_QUERIES,\n    \"global_score_threshold\": SCORE_THRESHOLD,\n    \"global_vision_model_question\": VISION_MODEL_QUESTION,\n    \"global_full_image_vision_model_question\": FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({\"global_monitor_index\": int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({\"global_text_queries\": x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({\"global_score_threshold\": float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({\"global_vision_model_question\": x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({\"global_full_image_vision_model_question\": x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    \"\"\"\n    Modifies the user input before it is processed by the LLM.\n    \"\"\"\n    return user_input\n\ndef output_modifier(output, state):\n    \"\"\"\n    Modifies the LLM output before it is presented in the UI.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the \"Autonomy_Tasks:\" trigger phrase in the LLM's output\n    if \"Autonomy_Tasks:\" in output:\n        # Clean the chat history\n        state[\"history\"] = history_modifier(state[\"history\"])\n        # Extract tasks and execute them\n        tasks_section = output.split(\"Autonomy_Tasks:\")[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n        with open(compressed_results_json_path, 'r') as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, 'w') as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"PageDown\":\n            pyautogui.press(\"pagedown\")\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n\ndef execute_oob_tasks_with_delay():\n    \"\"\"\n    Executes the saved OOB tasks after a delay.\n    \"\"\"\n    time.sleep(0.1)  # Short delay to ensure the LLM's response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"OOB_PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"OOB_PageDown\":\n            pyautogui.press(\"pagedown\")\n        else:\n            print(f\"Unknown OOB task: {task}\")\n\ndef delay_task(delay):\n    \"\"\"\n    Adds a delay in milliseconds.\n    \"\"\"\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    print(f\"Mouse moved to coordinates: ({x}, {y})\")\n\ndef left_click_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    \"\"\"\n    x, y = map(int, coordinates.split(\",\"))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f\"Mouse clicked at coordinates: ({x}, {y})\")\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot using the current values from the Gradio interface.\n    \"\"\"\n    take_screenshot(global_vars[\"global_monitor_index\"], global_vars[\"global_text_queries\"], global_vars[\"global_score_threshold\"], global_vars[\"global_vision_model_question\"], global_vars[\"global_full_image_vision_model_question\"])\n\ndef move_mouse_to_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef left_click_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    \"\"\"\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor's information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    \"\"\"\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f\"Mouse moved to coordinates: ({center_x}, {center_y})\")\n\ndef get_monitor_info(monitor_index):\n    \"\"\"\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor's information.\n    \"\"\"\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    \"\"\"\n    Modifies the chat history before the text generation in chat mode begins.\n    \"\"\"\n    # Clean the internal history\n    history[\"internal\"] = clean_chat_history(history[\"internal\"])\n    # Clean the visible history\n    history[\"visible\"] = clean_chat_history(history[\"visible\"])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n    with open(compressed_results_json_path, 'r') as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[\"internal\"].append([\"\", f\"<START_COMPRESSED_RESULTS>\\\\\\\\\\\\\\\\n{latest_compressed_results}\\\\\\\\\\\\\\\\n<END_COMPRESSED_RESULTS>\"])\n\n    return history\n\ndef clean_chat_history(history):\n    \"\"\"\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) > 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f\"{START_BOOKEND}.*?{END_BOOKEND}\", \"\", history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    \"\"\"\n    Appends the compressed results to the output.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n    output += f\"\\\\\\\\\\\\\\\\n\\\\\\\\\\\\\\\\n{START_BOOKEND}\\\\\\\\\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\\\\\\\\\n{END_BOOKEND}\"\n    return output\n\ndef text_input(text):\n    \"\"\"\n    Types the given text, replacing '/n' with newline characters.\n    \"\"\"\n    text = text.replace(\"/n\", \"\\\\\\\\n\")\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    \"\"\"\n    Presses the specified special key.\n    \"\"\"\n    if key == \"enter\":\n        pyautogui.press(\"enter\")\n    elif key == \"space\":\n        pyautogui.press(\"space\")\n    elif key == \"backspace\":\n        pyautogui.press(\"backspace\")\n    elif key == \"tab\":\n        pyautogui.press(\"tab\")\n    elif key == \"escape\":\n        pyautogui.press(\"esc\")\n    elif key == \"shift\":\n        pyautogui.press(\"shift\")\n    elif key == \"ctrl\":\n        pyautogui.press(\"ctrl\")\n    elif key == \"alt\":\n        pyautogui.press(\"alt\")\n    else:\n        print(f\"Unsupported special key: {key}\")\n\ndef function_key(key):\n    \"\"\"\n    Presses the specified function key.\n    \"\"\"\n    if key.startswith(\"f\") and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f\"Unsupported function key: {key}\")\n\ndef key_combo(combo):\n    \"\"\"\n    Presses the specified key combination.\n    \"\"\"\n    keys = combo.split(\"+\")\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyUp(key)\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()\n```\n\n### Explanation of Changes\n\n1. **Added `Human_annotated_image.png`**:\n   - The `query_image` function now creates two images: `annotated_image.png` (without red boxes) and `Human_annotated_image.png` (with red boxes).\n   - The `annotated_image.png` is used by the vision AI model, while the `Human_annotated_image.png` is for human inspection.\n\n2. **Updated `query_image` Function**:\n   - The function now returns three paths: `annotated_image_path`, `human_annotated_image_path`, and `json_output_path`.\n   - The red boxes are drawn only on the `human_img_pil` image, which is saved as `Human_annotated_image.png`.\n\n3. **Updated `take_screenshot` Function**:\n   - The function now calls `query_image` and receives the three paths.\n   - The `annotated_image_path` is used for processing with the vision AI model.\n\nWith these changes, you can now have a separate image for human inspection with the red boxes, while the vision AI model uses the image without the red boxes. This should help in reducing the obscurity caused by too many red lines."
        ],
        [
            "Awesome!! Wonderful :3\n\nOkay now for the other change we need to make.  Okay, I have two monitors a 1920x1080 (monitor 1) and a 384x2160 (monitor 0).\n\nNow I want the AI to have control of the smaller monitor 1, and it does so well.  However, I also want the AI to have the ability to click on the UI that it exists within (to send itself prompts specifically with the generate button), but to do this, I need to put the UI the AI is in, inside monitor 1.  This causes issues, one being there is less screen space for stuff like a web browser.\n\nSo I can take a screenshot of  monitor 0 with the UI and figure out which coordinates have the generate button on that monitor, but since the AI is operating in monitor 1, it can't use these coordinates because it tries to apply them to monitor 1.\n\nCan you change the code so if the coordinates are sent like this by the AI like this:\n\nMouseMove = 0,1278,2485\n\nThe AI will move the mouse to coordinates 1278,2485 on monitor 0 and the same for MouseClick:\n\nMouseClick = 0,1278,2485 \n\nThe mouse will click to those coordinates on Monitor 0.  And if it is 1,25,500 then it can moeve the mouse or mouse click to those coordinates on monitor 1.  So it goes Monitor,x,y",
            "Sure, I can help you with that. We'll modify the `move_mouse_to_coordinates_raw` and `left_click_coordinates_raw` functions to handle the monitor index along with the coordinates. This way, the AI can move the mouse and click on the specified monitor based on the provided coordinates.\n\nHere's the updated code with the requested changes:\n\n```python\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = \"extensions/Lucid_Autonomy/oob_tasks.json\"\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = \"clickable text hyperlinks,clickable icons,text bar field,clickable UI buttons,UI tabs\"  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.20  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = \"Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?\"\nFULL_IMAGE_VISION_MODEL_QUESTION = \"This is a computer screenshot, please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.  You are aiding in helping those with visual handicaps.\"\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(\"cuda\")\nelse:\n    device = torch.device(\"cpu\")\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = \"/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5\"\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(\"google/owlv2-base-patch16-ensemble\", trust_remote_code=True)\n    print(\"Owlv2 model loaded.\")\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(\"Owlv2 model unloaded.\")\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if \"int4\" in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 4-bit quantized model loaded on-demand.\")\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={\"\": device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(\"MiniCPM-Llama3 standard model loaded on-demand.\")\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(\"MiniCPM-Llama3 model unloaded.\")\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=\"screenshot.png\"):\n    \"\"\"\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is \"screenshot.png\").\n    \"\"\"\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = \"extensions/Lucid_Autonomy\"\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f\"Screenshot saved as {screenshot_path}\")\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, human_annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(annotated_image_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn't exist\n    output_dir = \"extensions/Lucid_Autonomy/ImageOutputTest\"\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f\"Failed to delete {file_path}. Reason: {e}\")\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(\"Processing complete. Check the output directory for the cropped images and results.\")\n\ndef query_image(img_path, text_queries, score_threshold):\n    \"\"\"\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    \"\"\"\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(\",\")\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=\"pt\").to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][\"boxes\"], results[0][\"scores\"], results[0][\"labels\"]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    human_img_pil = img.copy()\n    human_draw = ImageDraw.Draw(human_img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score < score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=\"red\")\n\n        # Draw the boxes for the human-annotated image\n        human_draw.rectangle(box, outline=\"red\", width=2)\n\n    # Save the annotated image for debugging\n    annotated_image_path = \"extensions/Lucid_Autonomy/annotated_image.png\"\n    img_pil.save(annotated_image_path)\n\n    # Save the human-annotated image for human inspection\n    human_annotated_image_path = \"extensions/Lucid_Autonomy/Human_annotated_image.png\"\n    human_img_pil.save(human_annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = \"extensions/Lucid_Autonomy/output.json\"\n    with open(json_output_path, 'w') as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, human_annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=\"extensions/Lucid_Autonomy/ImageOutputTest\", full_image_response=\"\"):\n    \"\"\"\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is \"extensions/Lucid_Autonomy/ImageOutputTest\").\n        full_image_response (str): The response from the vision model for the full image.\n    \"\"\"\n    with open(json_output_path, 'r') as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f\"{output_dir}/{label}_{i}.png\"\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            \"image_name\": f\"{label}_{i}.png\",\n            \"coordinates\": box,\n            \"description\": vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f\"{output_dir}/results.json\"\n    with open(results_json_path, 'w') as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f\"{output_dir}/compressedresults.json\"\n    compressed_results = [{\"image_name\": result[\"image_name\"], \"description\": result[\"description\"]} for result in results]\n    with open(compressed_results_json_path, 'w') as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            \"image_name\": \"full_image.png\",\n            \"coordinates\": [],\n            \"description\": full_image_response\n        })\n        compressed_results.append({\n            \"image_name\": \"full_image.png\",\n            \"description\": full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, 'w') as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, 'w') as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    \"\"\"\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    \"\"\"\n    image = Image.open(image_path).convert(\"RGB\")\n    messages = [\n        {\"role\": \"user\", \"content\": f\"{question}\"}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        \"max_new_tokens\": 896,\n        \"repetition_penalty\": 1.05,\n        \"num_beams\": 3,\n        \"top_p\": 0.8,\n        \"top_k\": 1,\n        \"temperature\": 0.7,\n        \"sampling\": True,\n    }\n    if \"int4\" in \"openbmb/MiniCPM-Llama3-V-2_5\":\n        # Disable streaming for the 4-bit model\n        generation_args[\"stream\"] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = \"\"\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end='')\n   \n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    \"global_monitor_index\": MONITOR_INDEX,\n    \"global_text_queries\": TEXT_QUERIES,\n    \"global_score_threshold\": SCORE_THRESHOLD,\n    \"global_vision_model_question\": VISION_MODEL_QUESTION,\n    \"global_full_image_vision_model_question\": FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    \"\"\"\n    Creates custom gradio elements when the UI is launched.\n    \"\"\"\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=\"Monitor Index\", value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=\"Text Queries\", value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=\"Score Threshold\", value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=\"Vision Model Question\", value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=\"Full Image Vision Model Question\", value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=\"Take Screenshot\")\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({\"global_monitor_index\": int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({\"global_text_queries\": x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({\"global_score_threshold\": float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({\"global_vision_model_question\": x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({\"global_full_image_vision_model_question\": x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    \"\"\"\n    Modifies the user input before it is processed by the LLM.\n    \"\"\"\n    return user_input\n\ndef output_modifier(output, state):\n    \"\"\"\n    Modifies the LLM output before it is presented in the UI.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the \"Autonomy_Tasks:\" trigger phrase in the LLM's output\n    if \"Autonomy_Tasks:\" in output:\n        # Clean the chat history\n        state[\"history\"] = history_modifier(state[\"history\"])\n        # Extract tasks and execute them\n        tasks_section = output.split(\"Autonomy_Tasks:\")[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n        with open(compressed_results_json_path, 'r') as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, 'w') as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    \"\"\"\n    Executes the tasks listed in the tasks_section.\n    \"\"\"\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(\"\\n\")\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(\"MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == \"TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"PageDown\":\n            pyautogui.press(\"pagedown\")\n        elif task.startswith(\"OOB_\"):\n            oob_tasks.append(task)\n        else:\n            print(f\"Unknown task: {task}\")\n\ndef execute_oob_tasks_with_delay():\n    \"\"\"\n    Executes the saved OOB tasks after a delay.\n    \"\"\"\n    time.sleep(0.1)  # Short delay to ensure the LLM's response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    \"\"\"\n    Executes the saved OOB tasks.\n    \"\"\"\n    with open(oob_tasks_file, 'r') as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(\"OOB_MouseMover =\"):\n            image_name = task.split(\"=\")[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            value = task.split(\"=\")[1].strip()\n            if \",\" in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(\"OOB_TextInput =\"):\n            text = task.split(\"=\")[1].strip()\n            text_input(text)\n        elif task.startswith(\"OOB_SpecialKey =\"):\n            key = task.split(\"=\")[1].strip()\n            special_key(key)\n        elif task.startswith(\"OOB_FunctionKey =\"):\n            key = task.split(\"=\")[1].strip()\n            function_key(key)\n        elif task.startswith(\"OOB_KeyCombo =\"):\n            combo = task.split(\"=\")[1].strip()\n            key_combo(combo)\n        elif task.startswith(\"OOB_Delay =\"):\n            delay = int(task.split(\"=\")[1].strip())\n            delay_task(delay)\n        elif task.startswith(\"OOB_MouseMove =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(\"OOB_MouseClick =\"):\n            coordinates = task.split(\"=\")[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == \"OOB_TakeScreenshot\":\n            take_screenshot_task()\n        elif task == \"OOB_PageUp\":\n            pyautogui.press(\"pageup\")\n        elif task == \"OOB_PageDown\":\n            pyautogui.press(\"pagedown\")\n        else:\n            print(f\"Unknown OOB task: {task}\")\n\ndef delay_task(delay):\n    \"\"\"\n    Adds a delay in milliseconds.\n    \"\"\"\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates on the specified monitor.\n\n    Args:\n        coordinates (str): A string containing the monitor index and coordinates in the format \"monitor_index,x,y\".\n    \"\"\"\n    monitor_index, x, y = map(int, coordinates.split(\",\"))\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    monitor = monitors[monitor_index]\n    monitor_x, monitor_y = monitor.x, monitor.y\n    absolute_x = monitor_x + x\n    absolute_y = monitor_y + y\n    pyautogui.moveTo(absolute_x, absolute_y)\n    print(f\"Mouse moved to coordinates: ({absolute_x}, {absolute_y}) on monitor {monitor_index}\")\n\ndef left_click_coordinates_raw(coordinates):\n    \"\"\"\n    Moves the mouse to the specified raw coordinates on the specified monitor and performs a left click.\n\n    Args:\n        coordinates (str): A string containing the monitor index and coordinates in the format \"monitor_index,x,y\".\n    \"\"\"\n    monitor_index, x, y = map(int, coordinates.split(\",\"))\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    monitor = monitors[monitor_index]\n    monitor_x, monitor_y = monitor.x, monitor.y\n    absolute_x = monitor_x + x\n    absolute_y = monitor_y + y\n    pyautogui.moveTo(absolute_x, absolute_y)\n    pyautogui.click()\n    print(f\"Mouse clicked at coordinates: ({absolute_x}, {absolute_y}) on monitor {monitor_index}\")\n\ndef take_screenshot_task():\n    \"\"\"\n    Takes a screenshot using the current values from the Gradio interface.\n    \"\"\"\n    take_screenshot(global_vars[\"global_monitor_index\"], global_vars[\"global_text_queries\"], global_vars[\"global_score_threshold\"], global_vars[\"global_vision_model_question\"], global_vars[\"global_full_image_vision_model_question\"])\n\ndef move_mouse_to_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef left_click_image(image_name):\n    \"\"\"\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    \"\"\"\n    results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/results.json\"\n    with open(results_json_path, 'r') as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[\"image_name\"] == image_name:\n            coordinates = result[\"coordinates\"]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f\"Image name '{image_name}' not found in the results.\")\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    \"\"\"\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor's information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    \"\"\"\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f\"Mouse moved to coordinates: ({center_x}, {center_y})\")\n\ndef get_monitor_info(monitor_index):\n    \"\"\"\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor's information.\n    \"\"\"\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index >= len(monitors):\n        raise ValueError(f\"Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.\")\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    \"\"\"\n    Modifies the chat history before the text generation in chat mode begins.\n    \"\"\"\n    # Clean the internal history\n    history[\"internal\"] = clean_chat_history(history[\"internal\"])\n    # Clean the visible history\n    history[\"visible\"] = clean_chat_history(history[\"visible\"])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = \"extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json\"\n    with open(compressed_results_json_path, 'r') as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[\"internal\"].append([\"\", f\"<START_COMPRESSED_RESULTS>\\\\\\\\\\\\\\\\n{latest_compressed_results}\\\\\\\\\\\\\\\\n<END_COMPRESSED_RESULTS>\"])\n\n    return history\n\ndef clean_chat_history(history):\n    \"\"\"\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) > 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f\"{START_BOOKEND}.*?{END_BOOKEND}\", \"\", history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    \"\"\"\n    Appends the compressed results to the output.\n    \"\"\"\n    START_BOOKEND = \"<START_COMPRESSED_RESULTS>\"\n    END_BOOKEND = \"<END_COMPRESSED_RESULTS>\"\n    output += f\"\\\\\\\\\\\\\\\\n\\\\\\\\\\\\\\\\n{START_BOOKEND}\\\\\\\\\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\\\\\\\\\n{END_BOOKEND}\"\n    return output\n\ndef text_input(text):\n    \"\"\"\n    Types the given text, replacing '/n' with newline characters.\n    \"\"\"\n    text = text.replace(\"/n\", \"\\\\\\\\n\")\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    \"\"\"\n    Presses the specified special key.\n    \"\"\"\n    if key == \"enter\":\n        pyautogui.press(\"enter\")\n    elif key == \"space\":\n        pyautogui.press(\"space\")\n    elif key == \"backspace\":\n        pyautogui.press(\"backspace\")\n    elif key == \"tab\":\n        pyautogui.press(\"tab\")\n    elif key == \"escape\":\n        pyautogui.press(\"esc\")\n    elif key == \"shift\":\n        pyautogui.press(\"shift\")\n    elif key == \"ctrl\":\n        pyautogui.press(\"ctrl\")\n    elif key == \"alt\":\n        pyautogui.press(\"alt\")\n    else:\n        print(f\"Unsupported special key: {key}\")\n\ndef function_key(key):\n    \"\"\"\n    Presses the specified function key.\n    \"\"\"\n    if key.startswith(\"f\") and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f\"Unsupported function key: {key}\")\n\ndef key_combo(combo):\n    \"\"\"\n    Presses the specified key combination.\n    \"\"\"\n    keys = combo.split(\"+\")\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [\"ctrl\", \"alt\", \"shift\"]:\n            pyautogui.keyUp(key)\n\nif __name__ == \"__main__\":\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=\"ui\"\n    ).launch()\n```\n\n### Explanation of Changes\n\n1. **Updated `move_mouse_to_coordinates_raw` and `left_click_coordinates_raw` Functions**:\n   - These functions now take a string containing the monitor index and coordinates in the format \"monitor_index,x,y\".\n   - The functions parse the monitor index and coordinates, calculate the absolute coordinates based on the specified monitor, and move the mouse or click at the specified coordinates.\n\n2. **Updated `execute_tasks` and `execute_oob_tasks` Functions**:\n   - These functions now handle the new format for `MouseMove` and `MouseClick` tasks, which include the monitor index.\n\nWith these changes, the AI can now move the mouse and click on the specified monitor based on the provided coordinates. This should allow the AI to interact with the UI on monitor 0 while operating on monitor 1."
        ]
    ],
    "visible": [
        [
            "",
            "How can I help you today?"
        ],
        [
            "Hello, how are you functioning today?",
            "I&#x27;m functioning well, thank you for asking! I&#x27;m here and ready to assist you. How about you? How can I make your day better?"
        ],
        [
            "Okay so in a separate conversation you have helped me write some code for a project.  I&#x27;ve had you write a summary from that conversation and I&#x27;ll provide it here:\n\n### Summary of Lucid_Autonomy Extension\n\n#### Introduction to Textgen\n\nTextgen is a powerful web-based interface for running large language models (LLMs) locally. It allows users to interact with various LLMs, providing a flexible and customizable environment for text generation tasks. The platform supports a wide range of models and offers numerous extensions to enhance its functionality.\n\n#### Lucid_Autonomy Extension Overview\n\nThe Lucid_Autonomy extension is designed to enable an LLM to interact autonomously with a computer using mouse and keyboard actions. This extension is part of the textgen ecosystem and aims to provide the LLM with &quot;eyes&quot; and &quot;hands&quot; to perform tasks such as moving the mouse to specific locations on the screen, clicking on elements, entering text, and more.\n\n#### Current Operation of the Extension\n\n1. **Initial Setup and Image Processing**:\n   - The extension takes a screenshot of a specified monitor in a multi-monitor setup.\n   - The screenshot is processed using the Owlv2 model to detect objects based on text queries.\n   - The detected objects are cropped and saved to the `ImageOutputTest` directory.\n   - The cropped images are processed with the MiniCPM-Llama3-V-2_5 model to generate descriptions.\n   - The results are saved in a `results.json` file, containing the image name, coordinates, and description.\n   - A `compressedresults.json` file is created, omitting the coordinates and containing only the image name and description.\n\n2. **LLM Interaction**:\n   - The LLM can interact with the extension using the trigger phrase &quot;Autonomy_Tasks:&quot;.\n   - If the LLM sends the trigger phrase without any instructions, the extension appends the contents of the `compressedresults.json` file to the LLM&#x27;s response.\n   - If the LLM sends the trigger phrase with instructions (e.g., `MouseMover = clickable icons_8.png`), the extension executes the specified tasks and then appends the `compressedresults.json` contents to the LLM&#x27;s response.\n\n3. **Task Execution**:\n   - The extension can handle various tasks such as moving the mouse to specific locations, clicking on elements, entering text, pressing special keys, function keys, key combinations, adding delays, and taking screenshots.\n   - The extension also supports &quot;Out of Body&quot; (OOB) commands, which are executed after the LLM&#x27;s response is complete. These commands are saved to a JSON file and executed with a short delay.\n\n4. **Full Image Processing**:\n   - The extension now processes the entire screenshot with the MiniCPM-Llama3-V-2_5 model using a separate question (`FULL_IMAGE_VISION_MODEL_QUESTION`).\n   - The response from the full image processing is added to the `results.json` and `compressedresults.json` files.\n\n#### Ultimate Goal of the Extension\n\nThe ultimate goal of the Lucid_Autonomy extension is to enable a large language model (LLM) to interact autonomously with a computer. The LLM can perform tasks such as moving the mouse to specific locations on the screen, clicking on elements, entering text, and more. The extension allows the LLM to operate a computer autonomously, effectively giving the LLM &quot;eyes&quot; and &quot;hands&quot; to interact with the user interface.\n\n\n#### Conclusion\n\nThe Lucid_Autonomy extension is a powerful tool that enables an LLM to interact autonomously with a computer. The current implementation allows the LLM to take screenshots, process images, and move the mouse to specific locations. The ultimate goal is to expand the extension&#x27;s capabilities to support a wide range of mouse and keyboard interactions, enabling the LLM to perform complex tasks autonomously.\n",
            "Thank you for providing the summary. It&#x27;s clear and comprehensive. To move forward, how would you like me to assist you with the Lucid_Autonomy extension? Are there specific features you&#x27;d like to add, modifications you&#x27;d like to make, or perhaps you need help with troubleshooting or optimizing the current implementation? Please let me know how I can best support you."
        ],
        [
            "Let me give you all the code, and then I want you to give me a summary detailing how the code functions and a list of all the current tools at the LLM&#x27;s disposal:\n\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = &quot;extensions/Lucid_Autonomy/oob_tasks.json&quot;\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = &quot;clickable hyperlinks,clickable icons,text bar field&quot;  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.2  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = &quot;Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?&quot;\nFULL_IMAGE_VISION_MODEL_QUESTION = &quot;Please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.&quot;\n\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(&quot;cuda&quot;)\nelse:\n    device = torch.device(&quot;cpu&quot;)\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = &quot;/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5&quot;\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True)\n    print(&quot;Owlv2 model loaded.&quot;)\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(&quot;Owlv2 model unloaded.&quot;)\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if &quot;int4&quot; in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 4-bit quantized model loaded on-demand.&quot;)\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 standard model loaded on-demand.&quot;)\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(&quot;MiniCPM-Llama3 model unloaded.&quot;)\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=&quot;screenshot.png&quot;):\n    &quot;&quot;&quot;\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is &quot;screenshot.png&quot;).\n    &quot;&quot;&quot;\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = &quot;extensions/Lucid_Autonomy&quot;\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f&quot;Screenshot saved as {screenshot_path}&quot;)\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn&#x27;t exist\n    output_dir = &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f&quot;Failed to delete {file_path}. Reason: {e}&quot;)\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(&quot;Processing complete. Check the output directory for the cropped images and results.&quot;)\n\n\n\ndef query_image(img_path, text_queries, score_threshold):\n    &quot;&quot;&quot;\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    &quot;&quot;&quot;\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(&quot;,&quot;)\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=&quot;pt&quot;).to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][&quot;boxes&quot;], results[0][&quot;scores&quot;], results[0][&quot;labels&quot;]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score &lt; score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        draw.rectangle(box, outline=&quot;red&quot;, width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=&quot;red&quot;)\n\n    # Save the annotated image for debugging\n    annotated_image_path = &quot;extensions/Lucid_Autonomy/annotated_image.png&quot;\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = &quot;extensions/Lucid_Autonomy/output.json&quot;\n    with open(json_output_path, &#x27;w&#x27;) as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=&quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;, full_image_response=&quot;&quot;):\n    &quot;&quot;&quot;\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;).\n        full_image_response (str): The response from the vision model for the full image.\n    &quot;&quot;&quot;\n    with open(json_output_path, &#x27;r&#x27;) as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f&quot;{output_dir}/{label}_{i}.png&quot;\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            &quot;image_name&quot;: f&quot;{label}_{i}.png&quot;,\n            &quot;coordinates&quot;: box,\n            &quot;description&quot;: vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f&quot;{output_dir}/results.json&quot;\n    with open(results_json_path, &#x27;w&#x27;) as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f&quot;{output_dir}/compressedresults.json&quot;\n    compressed_results = [{&quot;image_name&quot;: result[&quot;image_name&quot;], &quot;description&quot;: result[&quot;description&quot;]} for result in results]\n    with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;coordinates&quot;: [],\n            &quot;description&quot;: full_image_response\n        })\n        compressed_results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;description&quot;: full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, &#x27;w&#x27;) as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n            json.dump(compressed_results, f, indent=4)\n\n\n\ndef process_with_vision_model(image_path, question):\n    &quot;&quot;&quot;\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    &quot;&quot;&quot;\n    image = Image.open(image_path).convert(&quot;RGB&quot;)\n    messages = [\n        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;{question}&quot;}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        &quot;max_new_tokens&quot;: 896,\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;num_beams&quot;: 3,\n        &quot;top_p&quot;: 0.8,\n        &quot;top_k&quot;: 1,\n        &quot;temperature&quot;: 0.7,\n        &quot;sampling&quot;: True,\n    }\n    if &quot;int4&quot; in &quot;openbmb/MiniCPM-Llama3-V-2_5&quot;:\n        # Disable streaming for the 4-bit model\n        generation_args[&quot;stream&quot;] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = &quot;&quot;\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end=&#x27;&#x27;)\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\n\ndef input_modifier(user_input, state):\n    &quot;&quot;&quot;\n    Modifies the user input before it is processed by the LLM.\n    &quot;&quot;&quot;\n    return user_input\n\ndef output_modifier(output, state):\n    &quot;&quot;&quot;\n    Modifies the LLM output before it is presented in the UI.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the &quot;Autonomy_Tasks:&quot; trigger phrase in the LLM&#x27;s output\n    if &quot;Autonomy_Tasks:&quot; in output:\n        # Clean the chat history\n        state[&quot;history&quot;] = history_modifier(state[&quot;history&quot;])\n        # Extract tasks and execute them\n        tasks_section = output.split(&quot;Autonomy_Tasks:&quot;)[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n        with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, &#x27;w&#x27;) as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\n\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n\n\n\ndef execute_oob_tasks_with_delay():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks after a delay.\n    &quot;&quot;&quot;\n    time.sleep(0.1)  # Short delay to ensure the LLM&#x27;s response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n\n\n\n\ndef delay_task(delay):\n    &quot;&quot;&quot;\n    Adds a delay in milliseconds.\n    &quot;&quot;&quot;\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    print(f&quot;Mouse moved to coordinates: ({x}, {y})&quot;)\n\ndef left_click_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f&quot;Mouse clicked at coordinates: ({x}, {y})&quot;)\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot.\n    &quot;&quot;&quot;\n    take_screenshot(MONITOR_INDEX, TEXT_QUERIES, SCORE_THRESHOLD, VISION_MODEL_QUESTION, FULL_IMAGE_VISION_MODEL_QUESTION)\n\n\n\ndef move_mouse_to_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef left_click_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor&#x27;s information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f&quot;Mouse moved to coordinates: ({center_x}, {center_y})&quot;)\n\ndef get_monitor_info(monitor_index):\n    &quot;&quot;&quot;\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor&#x27;s information.\n    &quot;&quot;&quot;\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    &quot;&quot;&quot;\n    Modifies the chat history before the text generation in chat mode begins.\n    &quot;&quot;&quot;\n    # Clean the internal history\n    history[&quot;internal&quot;] = clean_chat_history(history[&quot;internal&quot;])\n    # Clean the visible history\n    history[&quot;visible&quot;] = clean_chat_history(history[&quot;visible&quot;])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n    with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[&quot;internal&quot;].append([&quot;&quot;, f&quot;&lt;START_COMPRESSED_RESULTS&gt;\\\\\\\\n{latest_compressed_results}\\\\\\\\n&lt;END_COMPRESSED_RESULTS&gt;&quot;])\n\n    return history\n\ndef clean_chat_history(history):\n    &quot;&quot;&quot;\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) &gt; 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f&quot;{START_BOOKEND}.*?{END_BOOKEND}&quot;, &quot;&quot;, history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    &quot;&quot;&quot;\n    Appends the compressed results to the output.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n    output += f&quot;\\\\\\\\n\\\\\\\\n{START_BOOKEND}\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\n{END_BOOKEND}&quot;\n    return output\n\ndef text_input(text):\n    &quot;&quot;&quot;\n    Types the given text, replacing &#x27;/n&#x27; with newline characters.\n    &quot;&quot;&quot;\n    text = text.replace(&quot;/n&quot;, &quot;\\\\n&quot;)\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    &quot;&quot;&quot;\n    Presses the specified special key.\n    &quot;&quot;&quot;\n    if key == &quot;enter&quot;:\n        pyautogui.press(&quot;enter&quot;)\n    elif key == &quot;space&quot;:\n        pyautogui.press(&quot;space&quot;)\n    elif key == &quot;backspace&quot;:\n        pyautogui.press(&quot;backspace&quot;)\n    elif key == &quot;tab&quot;:\n        pyautogui.press(&quot;tab&quot;)\n    elif key == &quot;escape&quot;:\n        pyautogui.press(&quot;esc&quot;)\n    elif key == &quot;shift&quot;:\n        pyautogui.press(&quot;shift&quot;)\n    elif key == &quot;ctrl&quot;:\n        pyautogui.press(&quot;ctrl&quot;)\n    elif key == &quot;alt&quot;:\n        pyautogui.press(&quot;alt&quot;)\n    else:\n        print(f&quot;Unsupported special key: {key}&quot;)\n\ndef function_key(key):\n    &quot;&quot;&quot;\n    Presses the specified function key.\n    &quot;&quot;&quot;\n    if key.startswith(&quot;f&quot;) and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f&quot;Unsupported function key: {key}&quot;)\n\ndef key_combo(combo):\n    &quot;&quot;&quot;\n    Presses the specified key combination.\n    &quot;&quot;&quot;\n    keys = combo.split(&quot;+&quot;)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyUp(key)\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()\n",
            "Thank you for providing the code. Here&#x27;s a summary of how the code functions and a list of the current tools at the LLM&#x27;s disposal:\n\n### Code Functionality Summary\n\n1. **Imports and Configurations**:\n   - The code imports necessary libraries such as `gradio`, `json`, `pyautogui`, `screeninfo`, `transformers`, `PIL`, `torch`, `os`, `shutil`, `re`, `time`, and `threading`.\n   - Configurable variables include `MONITOR_INDEX`, `TEXT_QUERIES`, `SCORE_THRESHOLD`, `VISION_MODEL_QUESTION`, and `FULL_IMAGE_VISION_MODEL_QUESTION`.\n\n2. **Model Loading and Unloading**:\n   - Functions to load and unload the `Owlv2` model (`load_owlv2_model` and `unload_owlv2_model`).\n   - Functions to load and unload the `MiniCPM-Llama3-V-2_5` model (`load_minicpm_llama_model` and `unload_minicpm_llama_model`).\n\n3. **Screenshot and Image Processing**:\n   - `take_screenshot`: Takes a screenshot of a specified monitor, processes it using the `Owlv2` model to detect objects, and then processes the entire screenshot and cropped images using the `MiniCPM-Llama3-V-2_5` model.\n   - `query_image`: Processes the image using the `Owlv2` model to detect objects based on text queries.\n   - `crop_images`: Crops out the detected objects from the input image based on the bounding box coordinates and processes the cropped images using the `MiniCPM-Llama3-V-2_5` model.\n   - `process_with_vision_model`: Processes the image with the `MiniCPM-Llama3-V-2_5` model and returns the response.\n\n4. **User Interface**:\n   - `ui`: Creates custom Gradio elements for the user interface, allowing users to input parameters and take screenshots.\n\n5. **Input and Output Modifiers**:\n   - `input_modifier`: Modifies the user input before it is processed by the LLM.\n   - `output_modifier`: Modifies the LLM output before it is presented in the UI, handling tasks and appending compressed results.\n\n6. **Task Execution**:\n   - `execute_tasks`: Executes tasks listed in the tasks section, such as moving the mouse, clicking, typing text, pressing special keys, function keys, key combinations, adding delays, and taking screenshots.\n   - `execute_oob_tasks_with_delay` and `execute_oob_tasks`: Handle the execution of out-of-band (OOB) tasks after a delay.\n\n7. **Helper Functions**:\n   - Various helper functions for delaying tasks, moving the mouse, clicking, typing text, pressing special keys, function keys, key combinations, and getting monitor information.\n\n### Current Tools at the LLM&#x27;s Disposal\n\n1. **Mouse Actions**:\n   - `MouseMover`: Moves the mouse to the center of a specified image.\n   - `MouseClick`: Moves the mouse to specified coordinates or an image and performs a left click.\n   - `MouseMove`: Moves the mouse to specified raw coordinates.\n\n2. **Keyboard Actions**:\n   - `TextInput`: Types the given text, replacing &#x27;/n&#x27; with newline characters.\n   - `SpecialKey`: Presses specified special keys (e.g., enter, space, backspace, tab, escape, shift, ctrl, alt).\n   - `FunctionKey`: Presses specified function keys (e.g., F1, F2, etc.).\n   - `KeyCombo`: Presses specified key combinations (e.g., ctrl+c, alt+tab).\n\n3. **Delay**:\n   - `Delay`: Adds a delay in milliseconds.\n\n4. **Screenshot**:\n   - `TakeScreenshot`: Takes a screenshot of the specified monitor.\n\n5. **Out-of-Body (OOB) Tasks**:\n   - `OOB_MouseMover`, `OOB_MouseClick`, `OOB_TextInput`, `OOB_SpecialKey`, `OOB_FunctionKey`, `OOB_KeyCombo`, `OOB_Delay`, `OOB_MouseMove`, `OOB_TakeScreenshot`: Similar to the regular tasks but executed after the LLM&#x27;s response is complete.\n\nThis summary provides an overview of the code&#x27;s functionality and the tools available to the LLM for autonomous interaction with the computer. If you need further details or assistance with specific parts of the code, feel free to ask!"
        ],
        [
            "Okay so you can see that there are gradio elements with text fields I can changes.  These elements work when I as the user press the &quot;Take Screenshot&quot; button, but when the LLM sends a command like this:\n\n```\nAutonomy_Tasks:\nTakeScreenshot\\\n```\n\nOnly the hard coded values from the extension code are used and not the information from the text fields when I update them.",
            "I understand the issue. The current implementation of the `take_screenshot` function uses hardcoded values for `MONITOR_INDEX`, `TEXT_QUERIES`, `SCORE_THRESHOLD`, `VISION_MODEL_QUESTION`, and `FULL_IMAGE_VISION_MODEL_QUESTION`. When the LLM sends a command like `TakeScreenshot`, it doesn&#x27;t take into account the updated values from the Gradio text fields.\n\nTo fix this, we need to modify the `take_screenshot_task` function to use the current values from the Gradio interface instead of the hardcoded values. Here&#x27;s how you can do it:\n\n1. **Store the Gradio interface inputs in global variables**: Update the global variables whenever the user changes the input fields in the Gradio interface.\n\n2. **Modify the `take_screenshot_task` function**: Make the `take_screenshot_task` function use these global variables instead of the hardcoded values.\n\nHere&#x27;s the updated code:\n\n```python\n# Global variables for Gradio inputs\nglobal_monitor_index = MONITOR_INDEX\nglobal_text_queries = TEXT_QUERIES\nglobal_score_threshold = SCORE_THRESHOLD\nglobal_vision_model_question = VISION_MODEL_QUESTION\nglobal_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: setattr(global_vars, &#x27;global_monitor_index&#x27;, int(x)), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: setattr(global_vars, &#x27;global_text_queries&#x27;, x), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: setattr(global_vars, &#x27;global_score_threshold&#x27;, float(x)), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: setattr(global_vars, &#x27;global_vision_model_question&#x27;, x), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: setattr(global_vars, &#x27;global_full_image_vision_model_question&#x27;, x), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot using the current values from the Gradio interface.\n    &quot;&quot;&quot;\n    take_screenshot(global_monitor_index, global_text_queries, global_score_threshold, global_vision_model_question, global_full_image_vision_model_question)\n\n# Define a class to hold global variables\nclass GlobalVars:\n    pass\n\nglobal_vars = GlobalVars()\nglobal_vars.global_monitor_index = MONITOR_INDEX\nglobal_vars.global_text_queries = TEXT_QUERIES\nglobal_vars.global_score_threshold = SCORE_THRESHOLD\nglobal_vars.global_vision_model_question = VISION_MODEL_QUESTION\nglobal_vars.global_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()\n```\n\nIn this updated code:\n- We define global variables to store the current values from the Gradio interface.\n- We update these global variables whenever the user changes the input fields in the Gradio interface.\n- The `take_screenshot_task` function now uses these global variables instead of the hardcoded values.\n\nThis should ensure that the `TakeScreenshot` command uses the updated values from the Gradio text fields."
        ],
        [
            "Unfortunately that did not work.  Can you look at the code in its current state, maybe I implemented your change incorrectly?  As a user using the UI I can press the &quot;Take Screenshot&quot; button and the changes I make to things like &quot;Score Threshold&quot; are overtly working (fewer icons are being picked up by owl).  But when I then ask the LLM to take a screenshot, it sends the right commands, a screenshot is taken, but I can tell the values in the UI that I changed are not making their way to the LLM&#x27;s utilization of &quot;Take Screenshot&quot; because there are a lot of icons recognized with the default value of 0.2.  So I know the LLM Autonomy_Task implementation of &quot;Take Screenshot&quot; is using the default value of 0.2 and not the changed value that I put in the UI.\n\nCurrent state of code:\n\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = &quot;extensions/Lucid_Autonomy/oob_tasks.json&quot;\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = &quot;clickable hyperlinks,clickable icons,text bar field&quot;  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.2  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = &quot;Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?&quot;\nFULL_IMAGE_VISION_MODEL_QUESTION = &quot;Please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.&quot;\n\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(&quot;cuda&quot;)\nelse:\n    device = torch.device(&quot;cpu&quot;)\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = &quot;/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5&quot;\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True)\n    print(&quot;Owlv2 model loaded.&quot;)\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(&quot;Owlv2 model unloaded.&quot;)\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if &quot;int4&quot; in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 4-bit quantized model loaded on-demand.&quot;)\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 standard model loaded on-demand.&quot;)\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(&quot;MiniCPM-Llama3 model unloaded.&quot;)\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=&quot;screenshot.png&quot;):\n    &quot;&quot;&quot;\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is &quot;screenshot.png&quot;).\n    &quot;&quot;&quot;\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = &quot;extensions/Lucid_Autonomy&quot;\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f&quot;Screenshot saved as {screenshot_path}&quot;)\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn&#x27;t exist\n    output_dir = &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f&quot;Failed to delete {file_path}. Reason: {e}&quot;)\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(&quot;Processing complete. Check the output directory for the cropped images and results.&quot;)\n\n\n\ndef query_image(img_path, text_queries, score_threshold):\n    &quot;&quot;&quot;\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    &quot;&quot;&quot;\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(&quot;,&quot;)\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=&quot;pt&quot;).to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][&quot;boxes&quot;], results[0][&quot;scores&quot;], results[0][&quot;labels&quot;]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score &lt; score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        draw.rectangle(box, outline=&quot;red&quot;, width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=&quot;red&quot;)\n\n    # Save the annotated image for debugging\n    annotated_image_path = &quot;extensions/Lucid_Autonomy/annotated_image.png&quot;\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = &quot;extensions/Lucid_Autonomy/output.json&quot;\n    with open(json_output_path, &#x27;w&#x27;) as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=&quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;, full_image_response=&quot;&quot;):\n    &quot;&quot;&quot;\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;).\n        full_image_response (str): The response from the vision model for the full image.\n    &quot;&quot;&quot;\n    with open(json_output_path, &#x27;r&#x27;) as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f&quot;{output_dir}/{label}_{i}.png&quot;\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            &quot;image_name&quot;: f&quot;{label}_{i}.png&quot;,\n            &quot;coordinates&quot;: box,\n            &quot;description&quot;: vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f&quot;{output_dir}/results.json&quot;\n    with open(results_json_path, &#x27;w&#x27;) as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f&quot;{output_dir}/compressedresults.json&quot;\n    compressed_results = [{&quot;image_name&quot;: result[&quot;image_name&quot;], &quot;description&quot;: result[&quot;description&quot;]} for result in results]\n    with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;coordinates&quot;: [],\n            &quot;description&quot;: full_image_response\n        })\n        compressed_results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;description&quot;: full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, &#x27;w&#x27;) as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n            json.dump(compressed_results, f, indent=4)\n\n\n\ndef process_with_vision_model(image_path, question):\n    &quot;&quot;&quot;\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    &quot;&quot;&quot;\n    image = Image.open(image_path).convert(&quot;RGB&quot;)\n    messages = [\n        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;{question}&quot;}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        &quot;max_new_tokens&quot;: 896,\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;num_beams&quot;: 3,\n        &quot;top_p&quot;: 0.8,\n        &quot;top_k&quot;: 1,\n        &quot;temperature&quot;: 0.7,\n        &quot;sampling&quot;: True,\n    }\n    if &quot;int4&quot; in &quot;openbmb/MiniCPM-Llama3-V-2_5&quot;:\n        # Disable streaming for the 4-bit model\n        generation_args[&quot;stream&quot;] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = &quot;&quot;\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end=&#x27;&#x27;)\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_monitor_index = MONITOR_INDEX\nglobal_text_queries = TEXT_QUERIES\nglobal_score_threshold = SCORE_THRESHOLD\nglobal_vision_model_question = VISION_MODEL_QUESTION\nglobal_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: setattr(global_vars, &#x27;global_monitor_index&#x27;, int(x)), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: setattr(global_vars, &#x27;global_text_queries&#x27;, x), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: setattr(global_vars, &#x27;global_score_threshold&#x27;, float(x)), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: setattr(global_vars, &#x27;global_vision_model_question&#x27;, x), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: setattr(global_vars, &#x27;global_full_image_vision_model_question&#x27;, x), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\n\ndef input_modifier(user_input, state):\n    &quot;&quot;&quot;\n    Modifies the user input before it is processed by the LLM.\n    &quot;&quot;&quot;\n    return user_input\n\ndef output_modifier(output, state):\n    &quot;&quot;&quot;\n    Modifies the LLM output before it is presented in the UI.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the &quot;Autonomy_Tasks:&quot; trigger phrase in the LLM&#x27;s output\n    if &quot;Autonomy_Tasks:&quot; in output:\n        # Clean the chat history\n        state[&quot;history&quot;] = history_modifier(state[&quot;history&quot;])\n        # Extract tasks and execute them\n        tasks_section = output.split(&quot;Autonomy_Tasks:&quot;)[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n        with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, &#x27;w&#x27;) as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\n\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n\n\n\ndef execute_oob_tasks_with_delay():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks after a delay.\n    &quot;&quot;&quot;\n    time.sleep(0.1)  # Short delay to ensure the LLM&#x27;s response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n\n\n\n\ndef delay_task(delay):\n    &quot;&quot;&quot;\n    Adds a delay in milliseconds.\n    &quot;&quot;&quot;\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    print(f&quot;Mouse moved to coordinates: ({x}, {y})&quot;)\n\ndef left_click_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f&quot;Mouse clicked at coordinates: ({x}, {y})&quot;)\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot using the current values from the Gradio interface.\n    &quot;&quot;&quot;\n    take_screenshot(global_monitor_index, global_text_queries, global_score_threshold, global_vision_model_question, global_full_image_vision_model_question)\n\n# Define a class to hold global variables\nclass GlobalVars:\n    pass\n\nglobal_vars = GlobalVars()\nglobal_vars.global_monitor_index = MONITOR_INDEX\nglobal_vars.global_text_queries = TEXT_QUERIES\nglobal_vars.global_score_threshold = SCORE_THRESHOLD\nglobal_vars.global_vision_model_question = VISION_MODEL_QUESTION\nglobal_vars.global_full_image_vision_model_question = FULL_IMAGE_VISION_MODEL_QUESTION\n\n\ndef move_mouse_to_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef left_click_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor&#x27;s information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f&quot;Mouse moved to coordinates: ({center_x}, {center_y})&quot;)\n\ndef get_monitor_info(monitor_index):\n    &quot;&quot;&quot;\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor&#x27;s information.\n    &quot;&quot;&quot;\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    &quot;&quot;&quot;\n    Modifies the chat history before the text generation in chat mode begins.\n    &quot;&quot;&quot;\n    # Clean the internal history\n    history[&quot;internal&quot;] = clean_chat_history(history[&quot;internal&quot;])\n    # Clean the visible history\n    history[&quot;visible&quot;] = clean_chat_history(history[&quot;visible&quot;])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n    with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[&quot;internal&quot;].append([&quot;&quot;, f&quot;&lt;START_COMPRESSED_RESULTS&gt;\\\\\\\\n{latest_compressed_results}\\\\\\\\n&lt;END_COMPRESSED_RESULTS&gt;&quot;])\n\n    return history\n\ndef clean_chat_history(history):\n    &quot;&quot;&quot;\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) &gt; 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f&quot;{START_BOOKEND}.*?{END_BOOKEND}&quot;, &quot;&quot;, history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    &quot;&quot;&quot;\n    Appends the compressed results to the output.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n    output += f&quot;\\\\\\\\n\\\\\\\\n{START_BOOKEND}\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\n{END_BOOKEND}&quot;\n    return output\n\ndef text_input(text):\n    &quot;&quot;&quot;\n    Types the given text, replacing &#x27;/n&#x27; with newline characters.\n    &quot;&quot;&quot;\n    text = text.replace(&quot;/n&quot;, &quot;\\\\n&quot;)\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    &quot;&quot;&quot;\n    Presses the specified special key.\n    &quot;&quot;&quot;\n    if key == &quot;enter&quot;:\n        pyautogui.press(&quot;enter&quot;)\n    elif key == &quot;space&quot;:\n        pyautogui.press(&quot;space&quot;)\n    elif key == &quot;backspace&quot;:\n        pyautogui.press(&quot;backspace&quot;)\n    elif key == &quot;tab&quot;:\n        pyautogui.press(&quot;tab&quot;)\n    elif key == &quot;escape&quot;:\n        pyautogui.press(&quot;esc&quot;)\n    elif key == &quot;shift&quot;:\n        pyautogui.press(&quot;shift&quot;)\n    elif key == &quot;ctrl&quot;:\n        pyautogui.press(&quot;ctrl&quot;)\n    elif key == &quot;alt&quot;:\n        pyautogui.press(&quot;alt&quot;)\n    else:\n        print(f&quot;Unsupported special key: {key}&quot;)\n\ndef function_key(key):\n    &quot;&quot;&quot;\n    Presses the specified function key.\n    &quot;&quot;&quot;\n    if key.startswith(&quot;f&quot;) and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f&quot;Unsupported function key: {key}&quot;)\n\ndef key_combo(combo):\n    &quot;&quot;&quot;\n    Presses the specified key combination.\n    &quot;&quot;&quot;\n    keys = combo.split(&quot;+&quot;)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyUp(key)\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()\n",
            "Thank you for providing the current state of your code. I see the issue now. The global variables are not being updated correctly because the `setattr` function is not being used properly. Additionally, the `GlobalVars` class is not being utilized correctly.\n\nLet&#x27;s correct the implementation to ensure that the global variables are updated properly when the user changes the input fields in the Gradio interface. Here&#x27;s the updated code:\n\n```python\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = &quot;extensions/Lucid_Autonomy/oob_tasks.json&quot;\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = &quot;clickable hyperlinks,clickable icons,text bar field&quot;  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.2  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = &quot;Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?&quot;\nFULL_IMAGE_VISION_MODEL_QUESTION = &quot;Please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.&quot;\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(&quot;cuda&quot;)\nelse:\n    device = torch.device(&quot;cpu&quot;)\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = &quot;/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5&quot;\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True)\n    print(&quot;Owlv2 model loaded.&quot;)\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(&quot;Owlv2 model unloaded.&quot;)\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if &quot;int4&quot; in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 4-bit quantized model loaded on-demand.&quot;)\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 standard model loaded on-demand.&quot;)\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(&quot;MiniCPM-Llama3 model unloaded.&quot;)\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=&quot;screenshot.png&quot;):\n    &quot;&quot;&quot;\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is &quot;screenshot.png&quot;).\n    &quot;&quot;&quot;\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = &quot;extensions/Lucid_Autonomy&quot;\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f&quot;Screenshot saved as {screenshot_path}&quot;)\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn&#x27;t exist\n    output_dir = &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f&quot;Failed to delete {file_path}. Reason: {e}&quot;)\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(&quot;Processing complete. Check the output directory for the cropped images and results.&quot;)\n\ndef query_image(img_path, text_queries, score_threshold):\n    &quot;&quot;&quot;\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    &quot;&quot;&quot;\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(&quot;,&quot;)\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=&quot;pt&quot;).to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][&quot;boxes&quot;], results[0][&quot;scores&quot;], results[0][&quot;labels&quot;]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score &lt; score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        draw.rectangle(box, outline=&quot;red&quot;, width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=&quot;red&quot;)\n\n    # Save the annotated image for debugging\n    annotated_image_path = &quot;extensions/Lucid_Autonomy/annotated_image.png&quot;\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = &quot;extensions/Lucid_Autonomy/output.json&quot;\n    with open(json_output_path, &#x27;w&#x27;) as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=&quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;, full_image_response=&quot;&quot;):\n    &quot;&quot;&quot;\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;).\n        full_image_response (str): The response from the vision model for the full image.\n    &quot;&quot;&quot;\n    with open(json_output_path, &#x27;r&#x27;) as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f&quot;{output_dir}/{label}_{i}.png&quot;\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            &quot;image_name&quot;: f&quot;{label}_{i}.png&quot;,\n            &quot;coordinates&quot;: box,\n            &quot;description&quot;: vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f&quot;{output_dir}/results.json&quot;\n    with open(results_json_path, &#x27;w&#x27;) as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f&quot;{output_dir}/compressedresults.json&quot;\n    compressed_results = [{&quot;image_name&quot;: result[&quot;image_name&quot;], &quot;description&quot;: result[&quot;description&quot;]} for result in results]\n    with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;coordinates&quot;: [],\n            &quot;description&quot;: full_image_response\n        })\n        compressed_results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;description&quot;: full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, &#x27;w&#x27;) as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    &quot;&quot;&quot;\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    &quot;&quot;&quot;\n    image = Image.open(image_path).convert(&quot;RGB&quot;)\n    messages = [\n        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;{question}&quot;}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        &quot;max_new_tokens&quot;: 896,\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;num_beams&quot;: 3,\n        &quot;top_p&quot;: 0.8,\n        &quot;top_k&quot;: 1,\n        &quot;temperature&quot;: 0.7,\n        &quot;sampling&quot;: True,\n    }\n    if &quot;int4&quot; in &quot;openbmb/MiniCPM-Llama3-V-2_5&quot;:\n        # Disable streaming for the 4-bit model\n        generation_args[&quot;stream&quot;] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = &quot;&quot;\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end=&#x27;&#x27;)\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    &quot;global_monitor_index&quot;: MONITOR_INDEX,\n    &quot;global_text_queries&quot;: TEXT_QUERIES,\n    &quot;global_score_threshold&quot;: SCORE_THRESHOLD,\n    &quot;global_vision_model_question&quot;: VISION_MODEL_QUESTION,\n    &quot;global_full_image_vision_model_question&quot;: FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({&quot;global_monitor_index&quot;: int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({&quot;global_text_queries&quot;: x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({&quot;global_score_threshold&quot;: float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({&quot;global_vision_model_question&quot;: x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({&quot;global_full_image_vision_model_question&quot;: x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    &quot;&quot;&quot;\n    Modifies the user input before it is processed by the LLM.\n    &quot;&quot;&quot;\n    return user_input\n\ndef output_modifier(output, state):\n    &quot;&quot;&quot;\n    Modifies the LLM output before it is presented in the UI.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the &quot;Autonomy_Tasks:&quot; trigger phrase in the LLM&#x27;s output\n    if &quot;Autonomy_Tasks:&quot; in output:\n        # Clean the chat history\n        state[&quot;history&quot;] = history_modifier(state[&quot;history&quot;])\n        # Extract tasks and execute them\n        tasks_section = output.split(&quot;Autonomy_Tasks:&quot;)[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n        with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, &#x27;w&#x27;) as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n\ndef execute_oob_tasks_with_delay():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks after a delay.\n    &quot;&quot;&quot;\n    time.sleep(0.1)  # Short delay to ensure the LLM&#x27;s response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n\ndef delay_task(delay):\n    &quot;&quot;&quot;\n    Adds a delay in milliseconds.\n    &quot;&quot;&quot;\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    print(f&quot;Mouse moved to coordinates: ({x}, {y})&quot;)\n\ndef left_click_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f&quot;Mouse clicked at coordinates: ({x}, {y})&quot;)\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot using the current values from the Gradio interface.\n    &quot;&quot;&quot;\n    take_screenshot(global_vars[&quot;global_monitor_index&quot;], global_vars[&quot;global_text_queries&quot;], global_vars[&quot;global_score_threshold&quot;], global_vars[&quot;global_vision_model_question&quot;], global_vars[&quot;global_full_image_vision_model_question&quot;])\n\ndef move_mouse_to_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef left_click_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor&#x27;s information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f&quot;Mouse moved to coordinates: ({center_x}, {center_y})&quot;)\n\ndef get_monitor_info(monitor_index):\n    &quot;&quot;&quot;\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor&#x27;s information.\n    &quot;&quot;&quot;\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    &quot;&quot;&quot;\n    Modifies the chat history before the text generation in chat mode begins.\n    &quot;&quot;&quot;\n    # Clean the internal history\n    history[&quot;internal&quot;] = clean_chat_history(history[&quot;internal&quot;])\n    # Clean the visible history\n    history[&quot;visible&quot;] = clean_chat_history(history[&quot;visible&quot;])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n    with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[&quot;internal&quot;].append([&quot;&quot;, f&quot;&lt;START_COMPRESSED_RESULTS&gt;\\\\\\\\n{latest_compressed_results}\\\\\\\\n&lt;END_COMPRESSED_RESULTS&gt;&quot;])\n\n    return history\n\ndef clean_chat_history(history):\n    &quot;&quot;&quot;\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) &gt; 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f&quot;{START_BOOKEND}.*?{END_BOOKEND}&quot;, &quot;&quot;, history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    &quot;&quot;&quot;\n    Appends the compressed results to the output.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n    output += f&quot;\\\\\\\\n\\\\\\\\n{START_BOOKEND}\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\n{END_BOOKEND}&quot;\n    return output\n\ndef text_input(text):\n    &quot;&quot;&quot;\n    Types the given text, replacing &#x27;/n&#x27; with newline characters.\n    &quot;&quot;&quot;\n    text = text.replace(&quot;/n&quot;, &quot;\\\\n&quot;)\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    &quot;&quot;&quot;\n    Presses the specified special key.\n    &quot;&quot;&quot;\n    if key == &quot;enter&quot;:\n        pyautogui.press(&quot;enter&quot;)\n    elif key == &quot;space&quot;:\n        pyautogui.press(&quot;space&quot;)\n    elif key == &quot;backspace&quot;:\n        pyautogui.press(&quot;backspace&quot;)\n    elif key == &quot;tab&quot;:\n        pyautogui.press(&quot;tab&quot;)\n    elif key == &quot;escape&quot;:\n        pyautogui.press(&quot;esc&quot;)\n    elif key == &quot;shift&quot;:\n        pyautogui.press(&quot;shift&quot;)\n    elif key == &quot;ctrl&quot;:\n        pyautogui.press(&quot;ctrl&quot;)\n    elif key == &quot;alt&quot;:\n        pyautogui.press(&quot;alt&quot;)\n    else:\n        print(f&quot;Unsupported special key: {key}&quot;)\n\ndef function_key(key):\n    &quot;&quot;&quot;\n    Presses the specified function key.\n    &quot;&quot;&quot;\n    if key.startswith(&quot;f&quot;) and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f&quot;Unsupported function key: {key}&quot;)\n\ndef key_combo(combo):\n    &quot;&quot;&quot;\n    Presses the specified key combination.\n    &quot;&quot;&quot;\n    keys = combo.split(&quot;+&quot;)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyUp(key)\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()\n```\n\nIn this updated code, I&#x27;ve corrected the implementation to ensure that the global variables are updated properly when the user changes the input fields in the Gradio interface. The `global_vars` dictionary is used to store the current values from the Gradio interface, and the `take_screenshot_task` function now uses these global variables instead of the hardcoded values. This should ensure that the `TakeScreenshot` command uses the updated values from the Gradio text fields."
        ],
        [
            "Yeass! Good job :3 *head pats*  okay let&#x27;s add one more function please then hopefully I can call it a night, I&#x27;m tired but this is so much fun!\n\nOkay this one is hopefully simple, can you add a tool that lets the LLM page up and page down? Don&#x27;t write any code, can you also explain the key combo section to me what does keyDown and keyUp mean in this context?",
            "I&#x27;m glad to hear that you&#x27;re enjoying this! Let&#x27;s add the functionality for the LLM to page up and page down.\n\n### Explanation of `keyDown` and `keyUp`\n\nIn the context of the `key_combo` function, `keyDown` and `keyUp` are used to simulate pressing and releasing keys on the keyboard. This is particularly useful for key combinations where you need to hold down one or more keys while pressing another key.\n\n- **`keyDown(key)`**: Simulates pressing down a key and holding it.\n- **`keyUp(key)`**: Simulates releasing a key that was previously pressed down.\n\nFor example, if you want to simulate pressing `Ctrl + C` (copy), you would:\n1. Press down the `Ctrl` key using `keyDown(&quot;ctrl&quot;)`.\n2. Press the `C` key using `pyautogui.press(&quot;c&quot;)`.\n3. Release the `Ctrl` key using `keyUp(&quot;ctrl&quot;)`.\n\nThis sequence ensures that the `Ctrl` key is held down while the `C` key is pressed, effectively performing the copy operation.\n\n### Adding Page Up and Page Down Functionality\n\nTo add the functionality for the LLM to page up and page down, you can extend the `execute_tasks` function to handle new tasks `PageUp` and `PageDown`. These tasks will use the `pyautogui.press` function to simulate pressing the `Page Up` and `Page Down` keys.\n\nHere&#x27;s how you can do it:\n\n1. **Extend the `execute_tasks` function**: Add new cases for `PageUp` and `PageDown`.\n2. **Update the `key_combo` function**: Ensure it can handle the new key combinations if needed.\n\n### Updated `execute_tasks` Function\n\n```python\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name"
        ],
        [
            "Sorry I had to stop you mid reply, but do we need to also update the execute_oob_tasks?  If so can you really review the working code closely to make sure you caught and contextualized everything that needs changing?",
            "Absolutely, let&#x27;s ensure that we update both the `execute_tasks` and `execute_oob_tasks` functions to handle the new `PageUp` and `PageDown` tasks. I&#x27;ll also review the working code closely to make sure everything is contextualized correctly.\n\n### Updated `execute_tasks` Function\n\n```python\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n```\n\n### Updated `execute_oob_tasks` Function\n\n```python\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;OOB_PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;OOB_PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n```\n\n### Explanation of `keyDown` and `keyUp`\n\nIn the context of the `key_combo` function, `keyDown` and `keyUp` are used to simulate pressing and releasing keys on the keyboard. This is particularly useful for key combinations where you need to hold down one or more keys while pressing another key.\n\n- **`keyDown(key)`**: Simulates pressing down a key and holding it.\n- **`keyUp(key)`**: Simulates releasing a key that was previously pressed down.\n\nFor example, if you want to simulate pressing `Ctrl + C` (copy), you would:\n1. Press down the `Ctrl` key using `keyDown(&quot;ctrl&quot;)`.\n2. Press the `C` key using `pyautogui.press(&quot;c&quot;)`.\n3. Release the `Ctrl` key using `keyUp(&quot;ctrl&quot;)`.\n\nThis sequence ensures that the `Ctrl` key is held down while the `C` key is pressed, effectively performing the copy operation.\n\n### Summary\n\nWith these updates, the LLM can now handle `PageUp` and `PageDown` tasks, both in the immediate execution and in the out-of-band (OOB) tasks. This should provide the functionality you need for paging up and down. If you have any further questions or need additional modifications, feel free to ask!"
        ],
        [
            "Okay, the code has been working perfectly for the last couple of days, I&#x27;ve been testing it out with a different model and it is surfing the internet and everything! :3 omg it&#x27;s amazing!  There are a few things I would like to change at this point however.\n\nOkay when the boxes are drawn on the screenshot and the annotated_image.png is created, sometimes there are too many red lines forming the boxes, which obscures the cropped out images for the minicpm vision model.\n\nSo as you&#x27;ll see in the code below, I&#x27;ve commented out the line that draws the boxes.  However, me being able to see the boxes is very useful.  So I would like you to keep the commented out red boxes (maybe I&#x27;ll turn that back on at some point) and keep the annoated_image.png going to the minicpm model as we have it, but I want you to create a new image called Human_annotated_image.png, I want it created in the same directory as the annotated_image.png.  This way I can look at the Human_annotated_image.png and the Vision AI can use the annotated_image.png\n\nI have an additional change after this, but it is more complex and we&#x27;ll just work on one at a time.\n\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = &quot;extensions/Lucid_Autonomy/oob_tasks.json&quot;\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = &quot;clickable text hyperlinks,clickable icons,text bar field,clickable UI buttons,UI tabs&quot;  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.20  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = &quot;Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?&quot;\nFULL_IMAGE_VISION_MODEL_QUESTION = &quot;This is a computer screenshot, please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.  You are aiding in helping those with visual handicaps.&quot;\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(&quot;cuda&quot;)\nelse:\n    device = torch.device(&quot;cpu&quot;)\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = &quot;/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5&quot;\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True)\n    print(&quot;Owlv2 model loaded.&quot;)\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(&quot;Owlv2 model unloaded.&quot;)\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if &quot;int4&quot; in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 4-bit quantized model loaded on-demand.&quot;)\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 standard model loaded on-demand.&quot;)\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(&quot;MiniCPM-Llama3 model unloaded.&quot;)\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=&quot;screenshot.png&quot;):\n    &quot;&quot;&quot;\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is &quot;screenshot.png&quot;).\n    &quot;&quot;&quot;\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = &quot;extensions/Lucid_Autonomy&quot;\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f&quot;Screenshot saved as {screenshot_path}&quot;)\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(screenshot_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn&#x27;t exist\n    output_dir = &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f&quot;Failed to delete {file_path}. Reason: {e}&quot;)\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(&quot;Processing complete. Check the output directory for the cropped images and results.&quot;)\n\ndef query_image(img_path, text_queries, score_threshold):\n    &quot;&quot;&quot;\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    &quot;&quot;&quot;\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(&quot;,&quot;)\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=&quot;pt&quot;).to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][&quot;boxes&quot;], results[0][&quot;scores&quot;], results[0][&quot;labels&quot;]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score &lt; score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        #draw.rectangle(box, outline=&quot;red&quot;, width=2)\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=&quot;red&quot;)\n\n    # Save the annotated image for debugging\n    annotated_image_path = &quot;extensions/Lucid_Autonomy/annotated_image.png&quot;\n    img_pil.save(annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = &quot;extensions/Lucid_Autonomy/output.json&quot;\n    with open(json_output_path, &#x27;w&#x27;) as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=&quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;, full_image_response=&quot;&quot;):\n    &quot;&quot;&quot;\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;).\n        full_image_response (str): The response from the vision model for the full image.\n    &quot;&quot;&quot;\n    with open(json_output_path, &#x27;r&#x27;) as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f&quot;{output_dir}/{label}_{i}.png&quot;\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            &quot;image_name&quot;: f&quot;{label}_{i}.png&quot;,\n            &quot;coordinates&quot;: box,\n            &quot;description&quot;: vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f&quot;{output_dir}/results.json&quot;\n    with open(results_json_path, &#x27;w&#x27;) as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f&quot;{output_dir}/compressedresults.json&quot;\n    compressed_results = [{&quot;image_name&quot;: result[&quot;image_name&quot;], &quot;description&quot;: result[&quot;description&quot;]} for result in results]\n    with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;coordinates&quot;: [],\n            &quot;description&quot;: full_image_response\n        })\n        compressed_results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;description&quot;: full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, &#x27;w&#x27;) as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    &quot;&quot;&quot;\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    &quot;&quot;&quot;\n    image = Image.open(image_path).convert(&quot;RGB&quot;)\n    messages = [\n        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;{question}&quot;}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        &quot;max_new_tokens&quot;: 896,\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;num_beams&quot;: 3,\n        &quot;top_p&quot;: 0.8,\n        &quot;top_k&quot;: 1,\n        &quot;temperature&quot;: 0.7,\n        &quot;sampling&quot;: True,\n    }\n    if &quot;int4&quot; in &quot;openbmb/MiniCPM-Llama3-V-2_5&quot;:\n        # Disable streaming for the 4-bit model\n        generation_args[&quot;stream&quot;] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = &quot;&quot;\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end=&#x27;&#x27;)\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    &quot;global_monitor_index&quot;: MONITOR_INDEX,\n    &quot;global_text_queries&quot;: TEXT_QUERIES,\n    &quot;global_score_threshold&quot;: SCORE_THRESHOLD,\n    &quot;global_vision_model_question&quot;: VISION_MODEL_QUESTION,\n    &quot;global_full_image_vision_model_question&quot;: FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({&quot;global_monitor_index&quot;: int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({&quot;global_text_queries&quot;: x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({&quot;global_score_threshold&quot;: float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({&quot;global_vision_model_question&quot;: x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({&quot;global_full_image_vision_model_question&quot;: x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    &quot;&quot;&quot;\n    Modifies the user input before it is processed by the LLM.\n    &quot;&quot;&quot;\n    return user_input\n\ndef output_modifier(output, state):\n    &quot;&quot;&quot;\n    Modifies the LLM output before it is presented in the UI.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the &quot;Autonomy_Tasks:&quot; trigger phrase in the LLM&#x27;s output\n    if &quot;Autonomy_Tasks:&quot; in output:\n        # Clean the chat history\n        state[&quot;history&quot;] = history_modifier(state[&quot;history&quot;])\n        # Extract tasks and execute them\n        tasks_section = output.split(&quot;Autonomy_Tasks:&quot;)[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n        with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, &#x27;w&#x27;) as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n\n\ndef execute_oob_tasks_with_delay():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks after a delay.\n    &quot;&quot;&quot;\n    time.sleep(0.1)  # Short delay to ensure the LLM&#x27;s response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;OOB_PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;OOB_PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n\n\ndef delay_task(delay):\n    &quot;&quot;&quot;\n    Adds a delay in milliseconds.\n    &quot;&quot;&quot;\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    print(f&quot;Mouse moved to coordinates: ({x}, {y})&quot;)\n\ndef left_click_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f&quot;Mouse clicked at coordinates: ({x}, {y})&quot;)\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot using the current values from the Gradio interface.\n    &quot;&quot;&quot;\n    take_screenshot(global_vars[&quot;global_monitor_index&quot;], global_vars[&quot;global_text_queries&quot;], global_vars[&quot;global_score_threshold&quot;], global_vars[&quot;global_vision_model_question&quot;], global_vars[&quot;global_full_image_vision_model_question&quot;])\n\ndef move_mouse_to_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef left_click_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor&#x27;s information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f&quot;Mouse moved to coordinates: ({center_x}, {center_y})&quot;)\n\ndef get_monitor_info(monitor_index):\n    &quot;&quot;&quot;\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor&#x27;s information.\n    &quot;&quot;&quot;\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    &quot;&quot;&quot;\n    Modifies the chat history before the text generation in chat mode begins.\n    &quot;&quot;&quot;\n    # Clean the internal history\n    history[&quot;internal&quot;] = clean_chat_history(history[&quot;internal&quot;])\n    # Clean the visible history\n    history[&quot;visible&quot;] = clean_chat_history(history[&quot;visible&quot;])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n    with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[&quot;internal&quot;].append([&quot;&quot;, f&quot;&lt;START_COMPRESSED_RESULTS&gt;\\\\\\\\\\\\\\\\n{latest_compressed_results}\\\\\\\\\\\\\\\\n&lt;END_COMPRESSED_RESULTS&gt;&quot;])\n\n    return history\n\ndef clean_chat_history(history):\n    &quot;&quot;&quot;\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) &gt; 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f&quot;{START_BOOKEND}.*?{END_BOOKEND}&quot;, &quot;&quot;, history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    &quot;&quot;&quot;\n    Appends the compressed results to the output.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n    output += f&quot;\\\\\\\\\\\\\\\\n\\\\\\\\\\\\\\\\n{START_BOOKEND}\\\\\\\\\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\\\\\\\\\n{END_BOOKEND}&quot;\n    return output\n\ndef text_input(text):\n    &quot;&quot;&quot;\n    Types the given text, replacing &#x27;/n&#x27; with newline characters.\n    &quot;&quot;&quot;\n    text = text.replace(&quot;/n&quot;, &quot;\\\\\\\\n&quot;)\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    &quot;&quot;&quot;\n    Presses the specified special key.\n    &quot;&quot;&quot;\n    if key == &quot;enter&quot;:\n        pyautogui.press(&quot;enter&quot;)\n    elif key == &quot;space&quot;:\n        pyautogui.press(&quot;space&quot;)\n    elif key == &quot;backspace&quot;:\n        pyautogui.press(&quot;backspace&quot;)\n    elif key == &quot;tab&quot;:\n        pyautogui.press(&quot;tab&quot;)\n    elif key == &quot;escape&quot;:\n        pyautogui.press(&quot;esc&quot;)\n    elif key == &quot;shift&quot;:\n        pyautogui.press(&quot;shift&quot;)\n    elif key == &quot;ctrl&quot;:\n        pyautogui.press(&quot;ctrl&quot;)\n    elif key == &quot;alt&quot;:\n        pyautogui.press(&quot;alt&quot;)\n    else:\n        print(f&quot;Unsupported special key: {key}&quot;)\n\ndef function_key(key):\n    &quot;&quot;&quot;\n    Presses the specified function key.\n    &quot;&quot;&quot;\n    if key.startswith(&quot;f&quot;) and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f&quot;Unsupported function key: {key}&quot;)\n\ndef key_combo(combo):\n    &quot;&quot;&quot;\n    Presses the specified key combination.\n    &quot;&quot;&quot;\n    keys = combo.split(&quot;+&quot;)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyUp(key)\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()",
            "Sure, let&#x27;s add the functionality to create a new image called `Human_annotated_image.png` that includes the red boxes for human inspection, while keeping the `annotated_image.png` without the red boxes for the vision AI model.\n\nHere&#x27;s the updated code with the requested changes:\n\n```python\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = &quot;extensions/Lucid_Autonomy/oob_tasks.json&quot;\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = &quot;clickable text hyperlinks,clickable icons,text bar field,clickable UI buttons,UI tabs&quot;  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.20  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = &quot;Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?&quot;\nFULL_IMAGE_VISION_MODEL_QUESTION = &quot;This is a computer screenshot, please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.  You are aiding in helping those with visual handicaps.&quot;\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(&quot;cuda&quot;)\nelse:\n    device = torch.device(&quot;cpu&quot;)\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = &quot;/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5&quot;\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True)\n    print(&quot;Owlv2 model loaded.&quot;)\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(&quot;Owlv2 model unloaded.&quot;)\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if &quot;int4&quot; in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 4-bit quantized model loaded on-demand.&quot;)\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 standard model loaded on-demand.&quot;)\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(&quot;MiniCPM-Llama3 model unloaded.&quot;)\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=&quot;screenshot.png&quot;):\n    &quot;&quot;&quot;\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is &quot;screenshot.png&quot;).\n    &quot;&quot;&quot;\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = &quot;extensions/Lucid_Autonomy&quot;\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f&quot;Screenshot saved as {screenshot_path}&quot;)\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, human_annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(annotated_image_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn&#x27;t exist\n    output_dir = &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f&quot;Failed to delete {file_path}. Reason: {e}&quot;)\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(&quot;Processing complete. Check the output directory for the cropped images and results.&quot;)\n\ndef query_image(img_path, text_queries, score_threshold):\n    &quot;&quot;&quot;\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    &quot;&quot;&quot;\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(&quot;,&quot;)\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=&quot;pt&quot;).to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][&quot;boxes&quot;], results[0][&quot;scores&quot;], results[0][&quot;labels&quot;]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    human_img_pil = img.copy()\n    human_draw = ImageDraw.Draw(human_img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score &lt; score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=&quot;red&quot;)\n\n        # Draw the boxes for the human-annotated image\n        human_draw.rectangle(box, outline=&quot;red&quot;, width=2)\n\n    # Save the annotated image for debugging\n    annotated_image_path = &quot;extensions/Lucid_Autonomy/annotated_image.png&quot;\n    img_pil.save(annotated_image_path)\n\n    # Save the human-annotated image for human inspection\n    human_annotated_image_path = &quot;extensions/Lucid_Autonomy/Human_annotated_image.png&quot;\n    human_img_pil.save(human_annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = &quot;extensions/Lucid_Autonomy/output.json&quot;\n    with open(json_output_path, &#x27;w&#x27;) as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, human_annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=&quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;, full_image_response=&quot;&quot;):\n    &quot;&quot;&quot;\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;).\n        full_image_response (str): The response from the vision model for the full image.\n    &quot;&quot;&quot;\n    with open(json_output_path, &#x27;r&#x27;) as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f&quot;{output_dir}/{label}_{i}.png&quot;\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            &quot;image_name&quot;: f&quot;{label}_{i}.png&quot;,\n            &quot;coordinates&quot;: box,\n            &quot;description&quot;: vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f&quot;{output_dir}/results.json&quot;\n    with open(results_json_path, &#x27;w&#x27;) as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f&quot;{output_dir}/compressedresults.json&quot;\n    compressed_results = [{&quot;image_name&quot;: result[&quot;image_name&quot;], &quot;description&quot;: result[&quot;description&quot;]} for result in results]\n    with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;coordinates&quot;: [],\n            &quot;description&quot;: full_image_response\n        })\n        compressed_results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;description&quot;: full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, &#x27;w&#x27;) as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    &quot;&quot;&quot;\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    &quot;&quot;&quot;\n    image = Image.open(image_path).convert(&quot;RGB&quot;)\n    messages = [\n        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;{question}&quot;}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        &quot;max_new_tokens&quot;: 896,\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;num_beams&quot;: 3,\n        &quot;top_p&quot;: 0.8,\n        &quot;top_k&quot;: 1,\n        &quot;temperature&quot;: 0.7,\n        &quot;sampling&quot;: True,\n    }\n    if &quot;int4&quot; in &quot;openbmb/MiniCPM-Llama3-V-2_5&quot;:\n        # Disable streaming for the 4-bit model\n        generation_args[&quot;stream&quot;] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = &quot;&quot;\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end=&#x27;&#x27;)\n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    &quot;global_monitor_index&quot;: MONITOR_INDEX,\n    &quot;global_text_queries&quot;: TEXT_QUERIES,\n    &quot;global_score_threshold&quot;: SCORE_THRESHOLD,\n    &quot;global_vision_model_question&quot;: VISION_MODEL_QUESTION,\n    &quot;global_full_image_vision_model_question&quot;: FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({&quot;global_monitor_index&quot;: int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({&quot;global_text_queries&quot;: x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({&quot;global_score_threshold&quot;: float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({&quot;global_vision_model_question&quot;: x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({&quot;global_full_image_vision_model_question&quot;: x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    &quot;&quot;&quot;\n    Modifies the user input before it is processed by the LLM.\n    &quot;&quot;&quot;\n    return user_input\n\ndef output_modifier(output, state):\n    &quot;&quot;&quot;\n    Modifies the LLM output before it is presented in the UI.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the &quot;Autonomy_Tasks:&quot; trigger phrase in the LLM&#x27;s output\n    if &quot;Autonomy_Tasks:&quot; in output:\n        # Clean the chat history\n        state[&quot;history&quot;] = history_modifier(state[&quot;history&quot;])\n        # Extract tasks and execute them\n        tasks_section = output.split(&quot;Autonomy_Tasks:&quot;)[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n        with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, &#x27;w&#x27;) as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n\ndef execute_oob_tasks_with_delay():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks after a delay.\n    &quot;&quot;&quot;\n    time.sleep(0.1)  # Short delay to ensure the LLM&#x27;s response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;OOB_PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;OOB_PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n\ndef delay_task(delay):\n    &quot;&quot;&quot;\n    Adds a delay in milliseconds.\n    &quot;&quot;&quot;\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    print(f&quot;Mouse moved to coordinates: ({x}, {y})&quot;)\n\ndef left_click_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates and performs a left click.\n    &quot;&quot;&quot;\n    x, y = map(int, coordinates.split(&quot;,&quot;))\n    pyautogui.moveTo(x, y)\n    pyautogui.click()\n    print(f&quot;Mouse clicked at coordinates: ({x}, {y})&quot;)\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot using the current values from the Gradio interface.\n    &quot;&quot;&quot;\n    take_screenshot(global_vars[&quot;global_monitor_index&quot;], global_vars[&quot;global_text_queries&quot;], global_vars[&quot;global_score_threshold&quot;], global_vars[&quot;global_vision_model_question&quot;], global_vars[&quot;global_full_image_vision_model_question&quot;])\n\ndef move_mouse_to_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef left_click_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor&#x27;s information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f&quot;Mouse moved to coordinates: ({center_x}, {center_y})&quot;)\n\ndef get_monitor_info(monitor_index):\n    &quot;&quot;&quot;\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor&#x27;s information.\n    &quot;&quot;&quot;\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    &quot;&quot;&quot;\n    Modifies the chat history before the text generation in chat mode begins.\n    &quot;&quot;&quot;\n    # Clean the internal history\n    history[&quot;internal&quot;] = clean_chat_history(history[&quot;internal&quot;])\n    # Clean the visible history\n    history[&quot;visible&quot;] = clean_chat_history(history[&quot;visible&quot;])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n    with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[&quot;internal&quot;].append([&quot;&quot;, f&quot;&lt;START_COMPRESSED_RESULTS&gt;\\\\\\\\\\\\\\\\n{latest_compressed_results}\\\\\\\\\\\\\\\\n&lt;END_COMPRESSED_RESULTS&gt;&quot;])\n\n    return history\n\ndef clean_chat_history(history):\n    &quot;&quot;&quot;\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) &gt; 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f&quot;{START_BOOKEND}.*?{END_BOOKEND}&quot;, &quot;&quot;, history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    &quot;&quot;&quot;\n    Appends the compressed results to the output.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n    output += f&quot;\\\\\\\\\\\\\\\\n\\\\\\\\\\\\\\\\n{START_BOOKEND}\\\\\\\\\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\\\\\\\\\n{END_BOOKEND}&quot;\n    return output\n\ndef text_input(text):\n    &quot;&quot;&quot;\n    Types the given text, replacing &#x27;/n&#x27; with newline characters.\n    &quot;&quot;&quot;\n    text = text.replace(&quot;/n&quot;, &quot;\\\\\\\\n&quot;)\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    &quot;&quot;&quot;\n    Presses the specified special key.\n    &quot;&quot;&quot;\n    if key == &quot;enter&quot;:\n        pyautogui.press(&quot;enter&quot;)\n    elif key == &quot;space&quot;:\n        pyautogui.press(&quot;space&quot;)\n    elif key == &quot;backspace&quot;:\n        pyautogui.press(&quot;backspace&quot;)\n    elif key == &quot;tab&quot;:\n        pyautogui.press(&quot;tab&quot;)\n    elif key == &quot;escape&quot;:\n        pyautogui.press(&quot;esc&quot;)\n    elif key == &quot;shift&quot;:\n        pyautogui.press(&quot;shift&quot;)\n    elif key == &quot;ctrl&quot;:\n        pyautogui.press(&quot;ctrl&quot;)\n    elif key == &quot;alt&quot;:\n        pyautogui.press(&quot;alt&quot;)\n    else:\n        print(f&quot;Unsupported special key: {key}&quot;)\n\ndef function_key(key):\n    &quot;&quot;&quot;\n    Presses the specified function key.\n    &quot;&quot;&quot;\n    if key.startswith(&quot;f&quot;) and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f&quot;Unsupported function key: {key}&quot;)\n\ndef key_combo(combo):\n    &quot;&quot;&quot;\n    Presses the specified key combination.\n    &quot;&quot;&quot;\n    keys = combo.split(&quot;+&quot;)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyUp(key)\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()\n```\n\n### Explanation of Changes\n\n1. **Added `Human_annotated_image.png`**:\n   - The `query_image` function now creates two images: `annotated_image.png` (without red boxes) and `Human_annotated_image.png` (with red boxes).\n   - The `annotated_image.png` is used by the vision AI model, while the `Human_annotated_image.png` is for human inspection.\n\n2. **Updated `query_image` Function**:\n   - The function now returns three paths: `annotated_image_path`, `human_annotated_image_path`, and `json_output_path`.\n   - The red boxes are drawn only on the `human_img_pil` image, which is saved as `Human_annotated_image.png`.\n\n3. **Updated `take_screenshot` Function**:\n   - The function now calls `query_image` and receives the three paths.\n   - The `annotated_image_path` is used for processing with the vision AI model.\n\nWith these changes, you can now have a separate image for human inspection with the red boxes, while the vision AI model uses the image without the red boxes. This should help in reducing the obscurity caused by too many red lines."
        ],
        [
            "Awesome!! Wonderful :3\n\nOkay now for the other change we need to make.  Okay, I have two monitors a 1920x1080 (monitor 1) and a 384x2160 (monitor 0).\n\nNow I want the AI to have control of the smaller monitor 1, and it does so well.  However, I also want the AI to have the ability to click on the UI that it exists within (to send itself prompts specifically with the generate button), but to do this, I need to put the UI the AI is in, inside monitor 1.  This causes issues, one being there is less screen space for stuff like a web browser.\n\nSo I can take a screenshot of  monitor 0 with the UI and figure out which coordinates have the generate button on that monitor, but since the AI is operating in monitor 1, it can&#x27;t use these coordinates because it tries to apply them to monitor 1.\n\nCan you change the code so if the coordinates are sent like this by the AI like this:\n\nMouseMove = 0,1278,2485\n\nThe AI will move the mouse to coordinates 1278,2485 on monitor 0 and the same for MouseClick:\n\nMouseClick = 0,1278,2485 \n\nThe mouse will click to those coordinates on Monitor 0.  And if it is 1,25,500 then it can moeve the mouse or mouse click to those coordinates on monitor 1.  So it goes Monitor,x,y",
            "Sure, I can help you with that. We&#x27;ll modify the `move_mouse_to_coordinates_raw` and `left_click_coordinates_raw` functions to handle the monitor index along with the coordinates. This way, the AI can move the mouse and click on the specified monitor based on the provided coordinates.\n\nHere&#x27;s the updated code with the requested changes:\n\n```python\nimport gradio as gr\nimport json\nimport pyautogui\nimport screeninfo\nfrom transformers import Owlv2Processor, Owlv2ForObjectDetection, AutoModel, AutoTokenizer\nfrom PIL import Image, ImageDraw\nimport torch\nimport os\nimport shutil\nimport re  # Import the re module for regular expression operations\nimport json\nimport time\nimport threading\n\noob_tasks_file = &quot;extensions/Lucid_Autonomy/oob_tasks.json&quot;\n\n# Configurable variables\nMONITOR_INDEX = 1  # Index of the monitor to capture (0 for the first monitor, 1 for the second monitor, etc.)\nTEXT_QUERIES = &quot;clickable text hyperlinks,clickable icons,text bar field,clickable UI buttons,UI tabs&quot;  # Comma-separated text queries for object detection\nSCORE_THRESHOLD = 0.20  # Score threshold for filtering low-probability predictions\nVISION_MODEL_QUESTION = &quot;Keep your response to less than 1 sentence: What is the text in this image or what does this image describe?&quot;\nFULL_IMAGE_VISION_MODEL_QUESTION = &quot;This is a computer screenshot, please describe what you see in this image, if you see UI elements describe them and their general location and if you see text read the text verbatim.  You are aiding in helping those with visual handicaps.&quot;\n\n# Use GPU if available\nif torch.cuda.is_available():\n    device = torch.device(&quot;cuda&quot;)\nelse:\n    device = torch.device(&quot;cpu&quot;)\n\n# Global variables for models\nowlv2_model = None\nowlv2_processor = None\nminicpm_llama_model = None\nminicpm_llama_tokenizer = None\n\nVISION_MODEL_ID = &quot;/home/myself/Desktop/miniCPM_llava3V/MiniCPM-Llama3-V-2_5&quot;\n\ndef load_owlv2_model():\n    global owlv2_model, owlv2_processor\n    owlv2_model = Owlv2ForObjectDetection.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True).to(device)\n    owlv2_processor = Owlv2Processor.from_pretrained(&quot;google/owlv2-base-patch16-ensemble&quot;, trust_remote_code=True)\n    print(&quot;Owlv2 model loaded.&quot;)\n\ndef unload_owlv2_model():\n    global owlv2_model, owlv2_processor\n    if owlv2_model is not None:\n        del owlv2_model\n        del owlv2_processor\n        torch.cuda.empty_cache()\n        print(&quot;Owlv2 model unloaded.&quot;)\n\ndef load_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if &quot;int4&quot; in VISION_MODEL_ID:\n        # Load the 4-bit quantized model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},\n            trust_remote_code=True,\n            torch_dtype=torch.float16  # Use float16 as per the example code for 4-bit models\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 4-bit quantized model loaded on-demand.&quot;)\n    else:\n        # Load the standard model and tokenizer\n        minicpm_llama_model = AutoModel.from_pretrained(\n            VISION_MODEL_ID,\n            device_map={&quot;&quot;: device},  # Use the specified CUDA device\n            trust_remote_code=True,\n            torch_dtype=torch.bfloat16\n        ).eval()\n        minicpm_llama_tokenizer = AutoTokenizer.from_pretrained(\n            VISION_MODEL_ID,\n            trust_remote_code=True\n        )\n        print(&quot;MiniCPM-Llama3 standard model loaded on-demand.&quot;)\n\ndef unload_minicpm_llama_model():\n    global minicpm_llama_model, minicpm_llama_tokenizer\n    if minicpm_llama_model is not None:\n        del minicpm_llama_model\n        del minicpm_llama_tokenizer\n        torch.cuda.empty_cache()\n        print(&quot;MiniCPM-Llama3 model unloaded.&quot;)\n\ndef take_screenshot(monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question, output_filename=&quot;screenshot.png&quot;):\n    &quot;&quot;&quot;\n    Takes a screenshot of a specific monitor in a multi-monitor setup.\n\n    Args:\n        monitor_index (int): The index of the monitor to capture.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n        vision_model_question (str): The question to ask the vision model for cropped images.\n        full_image_vision_model_question (str): The question to ask the vision model for the full image.\n        output_filename (str): The filename to save the screenshot as (default is &quot;screenshot.png&quot;).\n    &quot;&quot;&quot;\n    monitor_index = int(monitor_index)\n    score_threshold = float(score_threshold)\n\n    # Get information about all monitors\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    # Get the region of the specified monitor\n    monitor = monitors[monitor_index]\n    region = (monitor.x, monitor.y, monitor.width, monitor.height)\n\n    # Take the screenshot\n    output_dir = &quot;extensions/Lucid_Autonomy&quot;\n    os.makedirs(output_dir, exist_ok=True)\n    screenshot_path = os.path.join(output_dir, output_filename)\n    screenshot = pyautogui.screenshot(region=region)\n    screenshot.save(screenshot_path)\n    print(f&quot;Screenshot saved as {screenshot_path}&quot;)\n\n    # Process the image using the Owlv2 model\n    annotated_image_path, human_annotated_image_path, json_output_path = query_image(screenshot_path, text_queries, score_threshold)\n\n    # Load the miniCPM-lama3-v-2_5 model\n    load_minicpm_llama_model()\n\n    # Process the entire screenshot with the minicpm model\n    full_image_response = process_with_vision_model(annotated_image_path, full_image_vision_model_question)\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing the full image\n    unload_minicpm_llama_model()\n\n    # Create the output directory if it doesn&#x27;t exist\n    output_dir = &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;\n    os.makedirs(output_dir, exist_ok=True)\n\n    # Clear the output directory\n    for filename in os.listdir(output_dir):\n        file_path = os.path.join(output_dir, filename)\n        try:\n            if os.path.isfile(file_path) or os.path.islink(file_path):\n                os.unlink(file_path)\n            elif os.path.isdir(file_path):\n                shutil.rmtree(file_path)\n        except Exception as e:\n            print(f&quot;Failed to delete {file_path}. Reason: {e}&quot;)\n\n    # Crop the detected objects and save them to the output directory\n    crop_images(annotated_image_path, json_output_path, output_dir, full_image_response)\n\n    print(&quot;Processing complete. Check the output directory for the cropped images and results.&quot;)\n\ndef query_image(img_path, text_queries, score_threshold):\n    &quot;&quot;&quot;\n    Processes the image using the Owlv2 model to detect objects based on text queries.\n\n    Args:\n        img_path (str): The path to the input image.\n        text_queries (str): Comma-separated text queries for object detection.\n        score_threshold (float): The score threshold for filtering low-probability predictions.\n    &quot;&quot;&quot;\n    load_owlv2_model()\n\n    img = Image.open(img_path)\n    text_queries = text_queries.split(&quot;,&quot;)\n\n    size = max(img.size)\n    target_sizes = torch.Tensor([[size, size]])\n    inputs = owlv2_processor(text=text_queries, images=img, return_tensors=&quot;pt&quot;).to(device)\n\n    with torch.no_grad():\n        outputs = owlv2_model(**inputs)\n\n    outputs.logits = outputs.logits.cpu()\n    outputs.pred_boxes = outputs.pred_boxes.cpu()\n    results = owlv2_processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes)\n    boxes, scores, labels = results[0][&quot;boxes&quot;], results[0][&quot;scores&quot;], results[0][&quot;labels&quot;]\n\n    result_labels = []\n    img_pil = img.copy()\n    draw = ImageDraw.Draw(img_pil)\n\n    human_img_pil = img.copy()\n    human_draw = ImageDraw.Draw(human_img_pil)\n\n    for box, score, label in zip(boxes, scores, labels):\n        box = [int(i) for i in box.tolist()]\n        if score &lt; score_threshold:\n            continue\n        result_labels.append([box, text_queries[label.item()]])\n        # Do not draw the text label within the bounding box\n        # draw.text((box[0], box[1]), text_queries[label.item()], fill=&quot;red&quot;)\n\n        # Draw the boxes for the human-annotated image\n        human_draw.rectangle(box, outline=&quot;red&quot;, width=2)\n\n    # Save the annotated image for debugging\n    annotated_image_path = &quot;extensions/Lucid_Autonomy/annotated_image.png&quot;\n    img_pil.save(annotated_image_path)\n\n    # Save the human-annotated image for human inspection\n    human_annotated_image_path = &quot;extensions/Lucid_Autonomy/Human_annotated_image.png&quot;\n    human_img_pil.save(human_annotated_image_path)\n\n    # Save the JSON output to a file\n    json_output_path = &quot;extensions/Lucid_Autonomy/output.json&quot;\n    with open(json_output_path, &#x27;w&#x27;) as f:\n        json.dump(result_labels, f)\n\n    unload_owlv2_model()\n\n    return annotated_image_path, human_annotated_image_path, json_output_path\n\ndef crop_images(image_path, json_output_path, output_dir=&quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;, full_image_response=&quot;&quot;):\n    &quot;&quot;&quot;\n    Crops out the detected objects from the input image based on the bounding box coordinates.\n\n    Args:\n        image_path (str): The path to the input image.\n        json_output_path (str): The path to the JSON file containing the bounding box coordinates.\n        output_dir (str): The directory to save the cropped images (default is &quot;extensions/Lucid_Autonomy/ImageOutputTest&quot;).\n        full_image_response (str): The response from the vision model for the full image.\n    &quot;&quot;&quot;\n    with open(json_output_path, &#x27;r&#x27;) as f:\n        data = json.load(f)\n\n    img = Image.open(image_path)\n\n    results = []\n\n    # Load the miniCPM-lama3-v-2_5 model once before processing the images\n    load_minicpm_llama_model()\n\n    for i, item in enumerate(data):\n        box = item[0]\n        label = item[1]\n        cropped_img = img.crop(box)\n        cropped_img_path = f&quot;{output_dir}/{label}_{i}.png&quot;\n        cropped_img.save(cropped_img_path)\n\n        # Process the cropped image with the miniCPM-lama3-v-2_5 model\n        vision_model_response = process_with_vision_model(cropped_img_path, VISION_MODEL_QUESTION)\n\n        # Store the results in a structured format\n        results.append({\n            &quot;image_name&quot;: f&quot;{label}_{i}.png&quot;,\n            &quot;coordinates&quot;: box,\n            &quot;description&quot;: vision_model_response\n        })\n\n    # Unload the miniCPM-lama3-v-2_5 model after processing all images\n    unload_minicpm_llama_model()\n\n    # Save the results to a JSON file\n    results_json_path = f&quot;{output_dir}/results.json&quot;\n    with open(results_json_path, &#x27;w&#x27;) as f:\n        json.dump(results, f, indent=4)\n\n    # Create the compressed results JSON file\n    compressed_results_json_path = f&quot;{output_dir}/compressedresults.json&quot;\n    compressed_results = [{&quot;image_name&quot;: result[&quot;image_name&quot;], &quot;description&quot;: result[&quot;description&quot;]} for result in results]\n    with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n        json.dump(compressed_results, f, indent=4)\n\n    # Add the full image response to the results\n    if full_image_response:\n        results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;coordinates&quot;: [],\n            &quot;description&quot;: full_image_response\n        })\n        compressed_results.append({\n            &quot;image_name&quot;: &quot;full_image.png&quot;,\n            &quot;description&quot;: full_image_response\n        })\n\n        # Save the updated results to the JSON files\n        with open(results_json_path, &#x27;w&#x27;) as f:\n            json.dump(results, f, indent=4)\n        with open(compressed_results_json_path, &#x27;w&#x27;) as f:\n            json.dump(compressed_results, f, indent=4)\n\ndef process_with_vision_model(image_path, question):\n    &quot;&quot;&quot;\n    Processes the image with the miniCPM-lama3-v-2_5 model and returns the response.\n\n    Args:\n        image_path (str): The path to the input image.\n        question (str): The question to ask the vision model.\n\n    Returns:\n        str: The response from the vision model.\n    &quot;&quot;&quot;\n    image = Image.open(image_path).convert(&quot;RGB&quot;)\n    messages = [\n        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: f&quot;{question}&quot;}\n    ]\n    # Define the generation arguments\n    generation_args = {\n        &quot;max_new_tokens&quot;: 896,\n        &quot;repetition_penalty&quot;: 1.05,\n        &quot;num_beams&quot;: 3,\n        &quot;top_p&quot;: 0.8,\n        &quot;top_k&quot;: 1,\n        &quot;temperature&quot;: 0.7,\n        &quot;sampling&quot;: True,\n    }\n    if &quot;int4&quot; in &quot;openbmb/MiniCPM-Llama3-V-2_5&quot;:\n        # Disable streaming for the 4-bit model\n        generation_args[&quot;stream&quot;] = False\n        # Use the model.chat method with streaming enabled\n        vision_model_response = &quot;&quot;\n        for new_text in minicpm_llama_model.chat(\n                image=image,\n                msgs=messages,\n                tokenizer=minicpm_llama_tokenizer,\n                **generation_args\n        ):\n            vision_model_response += new_text\n            print(new_text, flush=True, end=&#x27;&#x27;)\n   \n    else:\n        minicpm_llama_model.to(device)\n        vision_model_response = minicpm_llama_model.chat(\n            image=image,\n            msgs=messages,\n            tokenizer=minicpm_llama_tokenizer,\n            **generation_args\n        )\n    return vision_model_response\n\n# Global variables for Gradio inputs\nglobal_vars = {\n    &quot;global_monitor_index&quot;: MONITOR_INDEX,\n    &quot;global_text_queries&quot;: TEXT_QUERIES,\n    &quot;global_score_threshold&quot;: SCORE_THRESHOLD,\n    &quot;global_vision_model_question&quot;: VISION_MODEL_QUESTION,\n    &quot;global_full_image_vision_model_question&quot;: FULL_IMAGE_VISION_MODEL_QUESTION\n}\n\ndef ui():\n    &quot;&quot;&quot;\n    Creates custom gradio elements when the UI is launched.\n    &quot;&quot;&quot;\n    with gr.Blocks() as demo:\n        with gr.Row():\n            with gr.Column():\n                monitor_index = gr.Textbox(label=&quot;Monitor Index&quot;, value=str(MONITOR_INDEX))\n                text_queries = gr.Textbox(label=&quot;Text Queries&quot;, value=TEXT_QUERIES)\n                score_threshold = gr.Textbox(label=&quot;Score Threshold&quot;, value=str(SCORE_THRESHOLD))\n                vision_model_question = gr.Textbox(label=&quot;Vision Model Question&quot;, value=VISION_MODEL_QUESTION)\n                full_image_vision_model_question = gr.Textbox(label=&quot;Full Image Vision Model Question&quot;, value=FULL_IMAGE_VISION_MODEL_QUESTION)\n                take_screenshot_button = gr.Button(value=&quot;Take Screenshot&quot;)\n\n        # Update global variables when the user changes the input fields\n        monitor_index.change(lambda x: global_vars.update({&quot;global_monitor_index&quot;: int(x)}), inputs=monitor_index, outputs=None)\n        text_queries.change(lambda x: global_vars.update({&quot;global_text_queries&quot;: x}), inputs=text_queries, outputs=None)\n        score_threshold.change(lambda x: global_vars.update({&quot;global_score_threshold&quot;: float(x)}), inputs=score_threshold, outputs=None)\n        vision_model_question.change(lambda x: global_vars.update({&quot;global_vision_model_question&quot;: x}), inputs=vision_model_question, outputs=None)\n        full_image_vision_model_question.change(lambda x: global_vars.update({&quot;global_full_image_vision_model_question&quot;: x}), inputs=full_image_vision_model_question, outputs=None)\n\n        take_screenshot_button.click(\n            fn=take_screenshot,\n            inputs=[monitor_index, text_queries, score_threshold, vision_model_question, full_image_vision_model_question],\n            outputs=None\n        )\n\n    return demo\n\ndef input_modifier(user_input, state):\n    &quot;&quot;&quot;\n    Modifies the user input before it is processed by the LLM.\n    &quot;&quot;&quot;\n    return user_input\n\ndef output_modifier(output, state):\n    &quot;&quot;&quot;\n    Modifies the LLM output before it is presented in the UI.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n\n    # Search for the &quot;Autonomy_Tasks:&quot; trigger phrase in the LLM&#x27;s output\n    if &quot;Autonomy_Tasks:&quot; in output:\n        # Clean the chat history\n        state[&quot;history&quot;] = history_modifier(state[&quot;history&quot;])\n        # Extract tasks and execute them\n        tasks_section = output.split(&quot;Autonomy_Tasks:&quot;)[-1].strip()\n        if tasks_section:\n            execute_tasks(tasks_section)\n        # Append new compressed results\n        compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n        with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n            compressed_results = json.load(f)\n        output = append_compressed_results(output, compressed_results)\n\n        # Write OOB tasks to a JSON file\n        with open(oob_tasks_file, &#x27;w&#x27;) as f:\n            json.dump(oob_tasks, f)\n\n        # Trigger the execution of OOB tasks after a delay\n        threading.Thread(target=execute_oob_tasks_with_delay).start()\n\n    return output\n\noob_tasks = []\n\ndef execute_tasks(tasks_section):\n    &quot;&quot;&quot;\n    Executes the tasks listed in the tasks_section.\n    &quot;&quot;&quot;\n    global oob_tasks\n    oob_tasks = []\n    tasks = tasks_section.split(&quot;\\n&quot;)\n    for task in tasks:\n        task = task.strip()\n        if task.startswith(&quot;MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task == &quot;TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        elif task.startswith(&quot;OOB_&quot;):\n            oob_tasks.append(task)\n        else:\n            print(f&quot;Unknown task: {task}&quot;)\n\ndef execute_oob_tasks_with_delay():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks after a delay.\n    &quot;&quot;&quot;\n    time.sleep(0.1)  # Short delay to ensure the LLM&#x27;s response has completed\n    execute_oob_tasks()\n\ndef execute_oob_tasks():\n    &quot;&quot;&quot;\n    Executes the saved OOB tasks.\n    &quot;&quot;&quot;\n    with open(oob_tasks_file, &#x27;r&#x27;) as f:\n        oob_tasks = json.load(f)\n\n    for task in oob_tasks:\n        task = task.strip()\n        if task.startswith(&quot;OOB_MouseMover =&quot;):\n            image_name = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_image(image_name)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            value = task.split(&quot;=&quot;)[1].strip()\n            if &quot;,&quot; in value:  # Check if the value contains coordinates\n                coordinates = value\n                left_click_coordinates_raw(coordinates)\n            else:  # Otherwise, treat it as an image name\n                image_name = value\n                left_click_image(image_name)\n        elif task.startswith(&quot;OOB_TextInput =&quot;):\n            text = task.split(&quot;=&quot;)[1].strip()\n            text_input(text)\n        elif task.startswith(&quot;OOB_SpecialKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            special_key(key)\n        elif task.startswith(&quot;OOB_FunctionKey =&quot;):\n            key = task.split(&quot;=&quot;)[1].strip()\n            function_key(key)\n        elif task.startswith(&quot;OOB_KeyCombo =&quot;):\n            combo = task.split(&quot;=&quot;)[1].strip()\n            key_combo(combo)\n        elif task.startswith(&quot;OOB_Delay =&quot;):\n            delay = int(task.split(&quot;=&quot;)[1].strip())\n            delay_task(delay)\n        elif task.startswith(&quot;OOB_MouseMove =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            move_mouse_to_coordinates_raw(coordinates)\n        elif task.startswith(&quot;OOB_MouseClick =&quot;):\n            coordinates = task.split(&quot;=&quot;)[1].strip()\n            left_click_coordinates_raw(coordinates)\n        elif task == &quot;OOB_TakeScreenshot&quot;:\n            take_screenshot_task()\n        elif task == &quot;OOB_PageUp&quot;:\n            pyautogui.press(&quot;pageup&quot;)\n        elif task == &quot;OOB_PageDown&quot;:\n            pyautogui.press(&quot;pagedown&quot;)\n        else:\n            print(f&quot;Unknown OOB task: {task}&quot;)\n\ndef delay_task(delay):\n    &quot;&quot;&quot;\n    Adds a delay in milliseconds.\n    &quot;&quot;&quot;\n    time.sleep(delay / 1000.0)\n\ndef move_mouse_to_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates on the specified monitor.\n\n    Args:\n        coordinates (str): A string containing the monitor index and coordinates in the format &quot;monitor_index,x,y&quot;.\n    &quot;&quot;&quot;\n    monitor_index, x, y = map(int, coordinates.split(&quot;,&quot;))\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    monitor = monitors[monitor_index]\n    monitor_x, monitor_y = monitor.x, monitor.y\n    absolute_x = monitor_x + x\n    absolute_y = monitor_y + y\n    pyautogui.moveTo(absolute_x, absolute_y)\n    print(f&quot;Mouse moved to coordinates: ({absolute_x}, {absolute_y}) on monitor {monitor_index}&quot;)\n\ndef left_click_coordinates_raw(coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse to the specified raw coordinates on the specified monitor and performs a left click.\n\n    Args:\n        coordinates (str): A string containing the monitor index and coordinates in the format &quot;monitor_index,x,y&quot;.\n    &quot;&quot;&quot;\n    monitor_index, x, y = map(int, coordinates.split(&quot;,&quot;))\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    monitor = monitors[monitor_index]\n    monitor_x, monitor_y = monitor.x, monitor.y\n    absolute_x = monitor_x + x\n    absolute_y = monitor_y + y\n    pyautogui.moveTo(absolute_x, absolute_y)\n    pyautogui.click()\n    print(f&quot;Mouse clicked at coordinates: ({absolute_x}, {absolute_y}) on monitor {monitor_index}&quot;)\n\ndef take_screenshot_task():\n    &quot;&quot;&quot;\n    Takes a screenshot using the current values from the Gradio interface.\n    &quot;&quot;&quot;\n    take_screenshot(global_vars[&quot;global_monitor_index&quot;], global_vars[&quot;global_text_queries&quot;], global_vars[&quot;global_score_threshold&quot;], global_vars[&quot;global_vision_model_question&quot;], global_vars[&quot;global_full_image_vision_model_question&quot;])\n\ndef move_mouse_to_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef left_click_image(image_name):\n    &quot;&quot;&quot;\n    Moves the mouse to the center of the box that defines the image and performs a left click.\n    &quot;&quot;&quot;\n    results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/results.json&quot;\n    with open(results_json_path, &#x27;r&#x27;) as f:\n        results = json.load(f)\n\n    for result in results:\n        if result[&quot;image_name&quot;] == image_name:\n            coordinates = result[&quot;coordinates&quot;]\n            monitor = get_monitor_info(MONITOR_INDEX)\n            move_mouse_to_coordinates(monitor, coordinates)\n            pyautogui.click()  # Perform a left click\n            break\n    else:\n        print(f&quot;Image name &#x27;{image_name}&#x27; not found in the results.&quot;)\n\ndef move_mouse_to_coordinates(monitor, coordinates):\n    &quot;&quot;&quot;\n    Moves the mouse cursor to the specified coordinates on the monitor.\n\n    Args:\n        monitor (dict): A dictionary containing the monitor&#x27;s information.\n        coordinates (list): A list containing the coordinates [x1, y1, x2, y2].\n    &quot;&quot;&quot;\n    x1, y1, x2, y2 = coordinates\n    monitor_x, monitor_y = monitor.x, monitor.y\n    center_x = monitor_x + (x1 + x2) // 2\n    center_y = monitor_y + (y1 + y2) // 2\n    pyautogui.moveTo(center_x, center_y)\n    print(f&quot;Mouse moved to coordinates: ({center_x}, {center_y})&quot;)\n\ndef get_monitor_info(monitor_index):\n    &quot;&quot;&quot;\n    Gets information about the specified monitor.\n\n    Args:\n        monitor_index (int): The index of the monitor to get information about.\n\n    Returns:\n        dict: A dictionary containing the monitor&#x27;s information.\n    &quot;&quot;&quot;\n    monitors = screeninfo.get_monitors()\n\n    if monitor_index &gt;= len(monitors):\n        raise ValueError(f&quot;Monitor index {monitor_index} is out of range. There are only {len(monitors)} monitors.&quot;)\n\n    return monitors[monitor_index]\n\ndef history_modifier(history):\n    &quot;&quot;&quot;\n    Modifies the chat history before the text generation in chat mode begins.\n    &quot;&quot;&quot;\n    # Clean the internal history\n    history[&quot;internal&quot;] = clean_chat_history(history[&quot;internal&quot;])\n    # Clean the visible history\n    history[&quot;visible&quot;] = clean_chat_history(history[&quot;visible&quot;])\n\n    # Inject the latest compressedresults.json into the internal history\n    compressed_results_json_path = &quot;extensions/Lucid_Autonomy/ImageOutputTest/compressedresults.json&quot;\n    with open(compressed_results_json_path, &#x27;r&#x27;) as f:\n        compressed_results = json.load(f)\n    latest_compressed_results = json.dumps(compressed_results, indent=4)\n    history[&quot;internal&quot;].append([&quot;&quot;, f&quot;&lt;START_COMPRESSED_RESULTS&gt;\\\\\\\\\\\\\\\\n{latest_compressed_results}\\\\\\\\\\\\\\\\n&lt;END_COMPRESSED_RESULTS&gt;&quot;])\n\n    return history\n\ndef clean_chat_history(history):\n    &quot;&quot;&quot;\n    Cleans the chat history by removing all but the latest instance of the contents between the bookends.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n\n    # Find all instances of the bookends\n    bookend_indices = []\n    for i, entry in enumerate(history):\n        if isinstance(entry, list) and len(entry) == 2:\n            if START_BOOKEND in entry[1] and END_BOOKEND in entry[1]:\n                bookend_indices.append(i)\n\n    # If there are multiple instances, remove all but the last one\n    if len(bookend_indices) &gt; 1:\n        for i in bookend_indices[:-1]:\n            history[i][1] = re.sub(f&quot;{START_BOOKEND}.*?{END_BOOKEND}&quot;, &quot;&quot;, history[i][1], flags=re.DOTALL)\n\n    return history\n\ndef append_compressed_results(output, compressed_results):\n    &quot;&quot;&quot;\n    Appends the compressed results to the output.\n    &quot;&quot;&quot;\n    START_BOOKEND = &quot;&lt;START_COMPRESSED_RESULTS&gt;&quot;\n    END_BOOKEND = &quot;&lt;END_COMPRESSED_RESULTS&gt;&quot;\n    output += f&quot;\\\\\\\\\\\\\\\\n\\\\\\\\\\\\\\\\n{START_BOOKEND}\\\\\\\\\\\\\\\\n{json.dumps(compressed_results, indent=4)}\\\\\\\\\\\\\\\\n{END_BOOKEND}&quot;\n    return output\n\ndef text_input(text):\n    &quot;&quot;&quot;\n    Types the given text, replacing &#x27;/n&#x27; with newline characters.\n    &quot;&quot;&quot;\n    text = text.replace(&quot;/n&quot;, &quot;\\\\\\\\n&quot;)\n    pyautogui.typewrite(text)\n\ndef special_key(key):\n    &quot;&quot;&quot;\n    Presses the specified special key.\n    &quot;&quot;&quot;\n    if key == &quot;enter&quot;:\n        pyautogui.press(&quot;enter&quot;)\n    elif key == &quot;space&quot;:\n        pyautogui.press(&quot;space&quot;)\n    elif key == &quot;backspace&quot;:\n        pyautogui.press(&quot;backspace&quot;)\n    elif key == &quot;tab&quot;:\n        pyautogui.press(&quot;tab&quot;)\n    elif key == &quot;escape&quot;:\n        pyautogui.press(&quot;esc&quot;)\n    elif key == &quot;shift&quot;:\n        pyautogui.press(&quot;shift&quot;)\n    elif key == &quot;ctrl&quot;:\n        pyautogui.press(&quot;ctrl&quot;)\n    elif key == &quot;alt&quot;:\n        pyautogui.press(&quot;alt&quot;)\n    else:\n        print(f&quot;Unsupported special key: {key}&quot;)\n\ndef function_key(key):\n    &quot;&quot;&quot;\n    Presses the specified function key.\n    &quot;&quot;&quot;\n    if key.startswith(&quot;f&quot;) and key[1:].isdigit():\n        pyautogui.press(key)\n    else:\n        print(f&quot;Unsupported function key: {key}&quot;)\n\ndef key_combo(combo):\n    &quot;&quot;&quot;\n    Presses the specified key combination.\n    &quot;&quot;&quot;\n    keys = combo.split(&quot;+&quot;)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyDown(key)\n        else:\n            pyautogui.press(key)\n    for key in keys:\n        if key in [&quot;ctrl&quot;, &quot;alt&quot;, &quot;shift&quot;]:\n            pyautogui.keyUp(key)\n\nif __name__ == &quot;__main__&quot;:\n    # Launch the Gradio interface with the specified UI components\n    gr.Interface(\n        fn=ui,\n        inputs=None,\n        outputs=&quot;ui&quot;\n    ).launch()\n```\n\n### Explanation of Changes\n\n1. **Updated `move_mouse_to_coordinates_raw` and `left_click_coordinates_raw` Functions**:\n   - These functions now take a string containing the monitor index and coordinates in the format &quot;monitor_index,x,y&quot;.\n   - The functions parse the monitor index and coordinates, calculate the absolute coordinates based on the specified monitor, and move the mouse or click at the specified coordinates.\n\n2. **Updated `execute_tasks` and `execute_oob_tasks` Functions**:\n   - These functions now handle the new format for `MouseMove` and `MouseClick` tasks, which include the monitor index.\n\nWith these changes, the AI can now move the mouse and click on the specified monitor based on the provided coordinates. This should allow the AI to interact with the UI on monitor 0 while operating on monitor 1."
        ]
    ]
}