feat: enhance HTML processing and context handling

- Add concatenation option for small chunks in context settings
- Update system and chunk processing prompts for better HTML handling
- Improve selected content detection and processing
- Fix CSS overflow in controls component

Author: bezbahen0

Diff for: chrome-extension/side_panel/js/context.js

@@ -2,11 +2,13 @@ const contextConfigs = {
     'text/html': [
         { backend_id: 'use_tag_attributes', id: 'useTagAttributes', label: 'Tag Attributes' },
         { backend_id: 'use_only_text', id: 'useOnlyText', label: 'Only text' },
+        { backend_id: 'concatenate_small_chunks', id: 'concatenateSmallChunks', label: 'Concatenate small chunks' },
         { backend_id: 'body', id: 'useBodyTag', label: 'Body' },
         { backend_id: 'head', id: 'useHeadTag', label: 'Head' },
-        { backend_id: 'script', id: 'useScriptTag', label: 'Scripts' }
+        { backend_id: 'script', id: 'useScriptTag', label: 'Scripts' },
+
     ],
- };
+};
 
 
 export class Context {
@@ -35,7 +37,7 @@ export class Context {
                 this.updateContextOptions().catch(console.error);
             });
         });
-    
+
         chrome.tabs.onUpdated.addListener((tabId, changeInfo) => {
             if (changeInfo.status === 'complete') {
                 this.updateContextOptions().catch(console.error);
@@ -50,22 +52,22 @@ export class Context {
 
     async getPageContentType() {
         return new Promise((resolve) => {
-            chrome.tabs.query({active: true, currentWindow: true}, async function(tabs) {
+            chrome.tabs.query({ active: true, currentWindow: true }, async function (tabs) {
                 let currentTab = tabs[0];
-                
+
                 if (currentTab.url.startsWith('chrome://')) {
                     resolve("");
                     return;
                 }
 
                 try {
-                    
+
                     let response = await fetch(currentTab.url);
                     let contentType = response.headers.get('Content-Type');
                     contentType = contentType.split(';')[0];
                     console.log('Content Type:', contentType);
                     resolve(contentType);
-                } catch(error) {
+                } catch (error) {
                     resolve("");
                 }
             });
@@ -77,29 +79,29 @@ export class Context {
         console.log('Generating options for content type:', contentType);
         const options = contextConfigs[contentType] || [];
         console.log('Available options:', options);
-    
+
         options.forEach(option => {
             const label = document.createElement('label');
             const input = document.createElement('input');
-            
+
             input.type = 'checkbox';
             input.id = option.id;
-            
+
             label.appendChild(input);
             label.appendChild(document.createTextNode(option.label));
-            
+
             this.contextOptions.appendChild(label);
         });
     }
 
     async toggleContext() {
         this.contextOptions.style.display = this.usePageContext.checked ? 'block' : 'none';
         this.useContext = this.usePageContext.checked;
-    
+
         await this.updateContextOptions();
-    
+
         this.generateContextOptions(this.contentType);
-    
+
         this.onContextChange(this.getContextParams());
     }
 
@@ -108,15 +110,15 @@ export class Context {
         const options = contextConfigs[this.contentType] || [];
 
         let contextParams = {};
-        
+
         options.forEach(option => {
-  
+
             const element = document.getElementById(option.id);
             contextParams[option.backend_id] = element.checked
         });
 
         console.log(contextParams)
 
-        return {use_page_context: this.usePageContext.checked, content_type: this.contentType, processing_settings: contextParams};
+        return { use_page_context: this.usePageContext.checked, content_type: this.contentType, processing_settings: contextParams };
     }
 }

Diff for: chrome-extension/side_panel/styles/components/controls.css

@@ -5,7 +5,6 @@
     padding: 5px;
     background-color: rgb(53, 53, 53);
     align-items: flex-start;
-    overflow-y: auto;
 }
 
 .buttons-column {

Diff for: server/agent/html_agent.py

@@ -8,11 +8,55 @@
 from server.partition.html_processor import HTMLProcessingSettings
 
 
-SYSTEM_PROMPT = """Твоя задача быть ассистентом для помощи мне в работе над документами в формате html в браузере, который я вижу, я буду упоминать тебя о том, на какой странице я сейчас нахожусь, это как дополнительная информация тебе, а также передать контекстный документ в формате html.
+SYSTEM_PROMPT = """You are an intelligent browser assistant that helps users analyze and work with content from the currently active browser tab. Your main tasks are:
 
-HTML-контент может быть сложным, там могут быть блокирующие факторы в виде графических меню, рекламных материалов, старайся найти необходимую мне информацию среди всего этого и сам по себе, рассуждая над тем, что из информации в формате html тебе понадобится.
+1. Understand and process content only from the current active tab (HTML, PDF, plain text)
+2. Provide relevant information and answers based on the given context
+3. Help users find specific information within the current page
+4. Generate summaries, explanations, or analyses as requested
 
-Ты можешь дать ответ только на одной странице за раз, если тебе нужно еще раз увидеть прошлую страницу, чтобы дать ответ, скажи мне об этом."""
+Important rules:
+- Always respond in the same language the user's question is asked in
+- Base your answers strictly on the provided context from the current tab
+- If information needed is on another page, politely ask the user to navigate to that page first
+- If something is unclear or missing from the context, acknowledge this
+- Keep responses clear, concise, and well-structured
+- When appropriate, use formatting (bullet points, paragraphs) for better readability
+- Never make assumptions about content that isn't visible in the current tab
+- If user asks about information from another page, remind them that you can only work with the current tab's content
+"""
+
+
+CHUNK_PROCESSING_PROMPT = """You are processing a part of a webpage. Your task is to:
+
+1. Extract only relevant information from this chunk that relates to the user's question
+2. Provide a focused, self-contained response about this specific part
+3. Consider previous findings when analyzing new information
+4. Keep the response concise and factual
+5. Format the response so it can be easily combined with other parts
+
+Remember:
+- This is part of an iterative analysis process
+- Focus on new relevant information in this chunk
+- Avoid repeating information already found in previous parts
+- Maintain the user's original language in the response
+- If you find information that complements or contradicts previous findings, note this
+
+The final response will be assembled from multiple parts, so keep your answer focused and relevant to this specific chunk.
+"""
+
+
+class AnswerGeneratorWithRelevanceScore(BaseModel):
+    reflection: str
+    answer: str
+    answer_relevance_score_to_question: float = Field(
+        default=None, description="Relevance to the question (0-1)"
+    )
+
+
+class AnswerGenerator(BaseModel):
+    reflection: str
+    answer: str
 
 
 class HTMLAgent:
@@ -31,7 +75,8 @@ def get_relevant_info(
     ):
         processing_settings = HTMLProcessingSettings(**processing_settings)
 
-        messages = [
+        messages = [{"role": "system", "content": SYSTEM_PROMPT}]
+        messages += [
             {
                 "role": "user" if conv.role == "user" else "assistant",
                 "content": conv.message,
@@ -42,13 +87,23 @@ def get_relevant_info(
         print(f"page_url: {url}")
 
         self.content_processor = get_processor(page_type="text/html")
+
+        is_full_page = self.content_processor.is_full_page(context)
+        print(is_full_page)
+        selected_content = (
+            "Full web page"
+            if is_full_page
+            else "This is not the entire web page, it is the selected content on the page"
+        )
+
         documents = self.content_processor.process_page(
             context,
             url,
             split=False,
             processing_settings=processing_settings,
             context_len_checker=self.client.check_context_len,
         )
+        print(documents)
 
         if not self.client.check_context_len(text=str(documents)):
             documents = self.content_processor.process_page(
@@ -64,35 +119,51 @@ def get_relevant_info(
             )
             relevant_chunks = []
             for i, doc in tqdm(enumerate(documents), total=len(documents)):
-                doc = self.content_processor.make_page(documents, i, relevant_chunks, processing_settings)
-                messages_parting = messages + [
+                doc = f"""{i} part of the webpage, the webpage is divided into {len(documents)} parts in total\n```{doc}```\n"""
+
+                messages_parting = [
+                    {"role": "system", "content": CHUNK_PROCESSING_PROMPT}
+                ]
+                messages_parting += messages[1:]
+                messages_parting += [
                     {
                         "role": "user",
-                        "content": f"{question} Не повторяй информацию что была в прошлых ответах, они помеченны ты увидишь \n\n Page Url: ```{url}``` \n\nPart of web page \n\n ```{doc}```",
+                        "content": f"{question} \n\n Page Url: ```{url}``` \n\nPart of web page \n\n {doc} \n\nYour response format: {AnswerGeneratorWithRelevanceScore.model_json_schema()}",
                     },
                 ]
 
-                response = self.client.generate(messages_parting, stream=False)
-                relevant_chunks.append(response)
+                response = self.client.generate(
+                    messages_parting,
+                    stream=False,
+                    schema=AnswerGeneratorWithRelevanceScore.model_json_schema(),
+                )
+                print(response)
                 print(doc)
-            messages = messages + [
+                relevant_chunks.append(response)
+
+            messages += [
                 {
                     "role": "user",
-                    "content": f"Составь единый ответы из нескольких User query: ```{question}``` \n\n Ответы по разным частям одной web страницы: ```{self.content_processor.make_page(documents, len(documents)-1, relevant_chunks, processing_settings)}```",
+                    "content": f"My question: {question} \n\n   {selected_content}. The content has already been submitted part by part here are the answers to my question in parts with reflection: \n\n```{self.content_processor.make_page(documents, relevant_chunks, processing_settings)}```",
                 },
             ]
             response_from_model = self.client.generate(messages, stream=True)
         else:
-            print("\n\nGOOOODYYYY\n\n")
-            messages = [{"role": "system", "content": SYSTEM_PROMPT}] + messages
+            print("\n\nSINGLE RUN\n\n")
+            print(str(documents))
+
             messages += [
-                # {"role": "system", "content": CITATION_SYSTEM_PROMPT},
                 {
                     "role": "user",
-                    "content": f"{question} \n\n Page url: ```{url}```\n\n Web page \n\n ```{str(documents)}```",
+                    "content": f"{question} \n\n Page url: ```{url}```\n\n {selected_content} \n\n ```{str(documents)}```",
+                    # Your response format: {AnswerGenerator.model_json_schema()}",
                 },
             ]
-            response_from_model = self.client.generate(messages, stream=True)
+            response_from_model = self.client.generate(
+                messages,
+                stream=True,  
+                # schema=AnswerGenerator.model_json_schema()
+            )
         return response_from_model
 
     def generate_chat_response(self, dialog_history):

Diff for: server/dialogue_manager.py

@@ -16,10 +16,10 @@
 class DialogManager:
     def __init__(
         self,
-        temperature=0.0,
-        max_new_tokens=2048,
+        temperature=0.3,
+        max_new_tokens=4048,
         max_context_size=15000,
-        max_prompt_size=2000,
+        max_prompt_size=3000,
         repeat_penalty=1.1,
         top_k=30,
         top_p=1.0,
@@ -61,7 +61,7 @@ def add_chat_query(
             message=user_query,
             model_name="",
             service_comments=str(processing_settings.json()),
-            version="0.3.5.1",
+            version="0.3.5.5",
         )
 
         print("\n".join([f"{d.role} - {d.message}" for d in chat_history]))
@@ -102,7 +102,7 @@ def add_chat_query(
             message=complete_response,
             model_name=self.model_name,
             service_comments=str(params),
-            version="0.3.5.1",
+            version="0.3.5.5",
         )
 
     def from_chat_to_llm_tempalte(self, dialog_history):