DH-5567/adding intermediate steps as the response (#440)

* DH-5567/adding intermediate steps as the response

* DH-5557add truncation

* Dh-5567/reformat

Author: MohammadrezaPourreza

dataherald/api/fastapi.py

@@ -731,7 +731,6 @@ def create_prompt_and_sql_generation(
             return error_response(
                 e, prompt_sql_generation_request.dict(), "sql_generation_not_created"
             )
-
         return SQLGenerationResponse(**sql_generation.dict())
 
     @override

dataherald/api/types/responses.py

@@ -5,7 +5,7 @@
 
 from dataherald.db_scanner.models.types import TableDescription
 from dataherald.sql_database.models.types import DatabaseConnection
-from dataherald.types import GoldenSQL, LLMConfig
+from dataherald.types import GoldenSQL, IntermediateStep, LLMConfig
 
 
 class BaseResponse(BaseModel):
@@ -33,6 +33,7 @@ class SQLGenerationResponse(BaseResponse):
     status: str
     completed_at: str | None
     llm_config: LLMConfig | None
+    intermediate_steps: list[IntermediateStep] | None
     sql: str | None
     tokens_used: int | None
     confidence_score: float | None

dataherald/services/sql_generations.py

@@ -48,6 +48,17 @@ def generate_response_with_timeout(self, sql_generator, user_prompt, db_connecti
             user_prompt=user_prompt, database_connection=db_connection
         )
 
+    def update_the_initial_sql_generation(
+        self, initial_sql_generation: SQLGeneration, sql_generation: SQLGeneration
+    ):
+        initial_sql_generation.sql = sql_generation.sql
+        initial_sql_generation.tokens_used = sql_generation.tokens_used
+        initial_sql_generation.completed_at = datetime.now()
+        initial_sql_generation.status = sql_generation.status
+        initial_sql_generation.error = sql_generation.error
+        initial_sql_generation.intermediate_steps = sql_generation.intermediate_steps
+        return self.sql_generation_repository.update(initial_sql_generation)
+
     def create(
         self, prompt_id: str, sql_generation_request: SQLGenerationRequest
     ) -> SQLGeneration:
@@ -153,12 +164,9 @@ def create(
             )
             initial_sql_generation.evaluate = sql_generation_request.evaluate
             initial_sql_generation.confidence_score = confidence_score
-        initial_sql_generation.sql = sql_generation.sql
-        initial_sql_generation.tokens_used = sql_generation.tokens_used
-        initial_sql_generation.completed_at = datetime.now()
-        initial_sql_generation.status = sql_generation.status
-        initial_sql_generation.error = sql_generation.error
-        return self.sql_generation_repository.update(initial_sql_generation)
+        return self.update_the_initial_sql_generation(
+            initial_sql_generation, sql_generation
+        )
 
     def start_streaming(
         self, prompt_id: str, sql_generation_request: SQLGenerationRequest, queue: Queue

dataherald/sql_generator/__init__.py

@@ -22,7 +22,7 @@
 from dataherald.sql_database.base import SQLDatabase, SQLInjectionError
 from dataherald.sql_database.models.types import DatabaseConnection
 from dataherald.sql_generator.create_sql_query_status import create_sql_query_status
-from dataherald.types import LLMConfig, Prompt, SQLGeneration
+from dataherald.types import IntermediateStep, LLMConfig, Prompt, SQLGeneration
 from dataherald.utils.strings import contains_line_breaks
 
 
@@ -74,20 +74,6 @@ def create_sql_query_status(
     ) -> SQLGeneration:
         return create_sql_query_status(db, query, sql_generation)
 
-    def format_intermediate_representations(
-        self, intermediate_representation: List[Tuple[AgentAction, str]]
-    ) -> List[str]:
-        """Formats the intermediate representation into a string."""
-        formatted_intermediate_representation = []
-        for item in intermediate_representation:
-            formatted_intermediate_representation.append(
-                f"Thought: '{str(item[0].log).split('Action:')[0]}'\n"
-                f"Action: '{item[0].tool}'\n"
-                f"Action Input: '{item[0].tool_input}'\n"
-                f"Observation: '{item[1]}'"
-            )
-        return formatted_intermediate_representation
-
     def format_sql_query(self, sql_query: str) -> str:
         comments = [
             match.group() for match in re.finditer(r"--.*$", sql_query, re.MULTILINE)
@@ -110,14 +96,53 @@ def extract_query_from_intermediate_steps(
             action = step[0]
             if type(action) == AgentAction and action.tool == "SqlDbQuery":
                 if "SELECT" in self.format_sql_query(action.tool_input).upper():
-                    sql_query = self.remove_markdown(sql_query)
+                    sql_query = self.remove_markdown(action.tool_input)
         if sql_query == "":
             for step in intermediate_steps:
                 action = step[0]
                 if "SELECT" in action.tool_input.upper():
-                    sql_query = self.remove_markdown(sql_query)
+                    sql_query = self.remove_markdown(action.tool_input)
+                    if not sql_query.upper().strip().startswith("SELECT"):
+                        sql_query = ""
         return sql_query
 
+    def construct_intermediate_steps(
+        self, intermediate_steps: List[Tuple[AgentAction, str]], suffix: str = ""
+    ) -> List[IntermediateStep]:
+        """Constructs the intermediate steps."""
+        formatted_intermediate_steps = []
+        for step in intermediate_steps:
+            if step[0].tool == "SqlDbQuery":
+                formatted_intermediate_steps.append(
+                    IntermediateStep(
+                        thought=str(step[0].log).split("Action:")[0],
+                        action=step[0].tool,
+                        action_input=step[0].tool_input,
+                        observation="QUERY RESULTS ARE NOT STORED FOR PRIVACY REASONS.",
+                    )
+                )
+            else:
+                formatted_intermediate_steps.append(
+                    IntermediateStep(
+                        thought=str(step[0].log).split("Action:")[0],
+                        action=step[0].tool,
+                        action_input=step[0].tool_input,
+                        observation=self.truncate_observations(step[1]),
+                    )
+                )
+        formatted_intermediate_steps[0].thought = suffix.split("Thought: ")[1].split(
+            "{agent_scratchpad}"
+        )[0]
+        return formatted_intermediate_steps
+
+    def truncate_observations(self, obervarion: str, max_length: int = 2000) -> str:
+        """Truncate the tool input."""
+        return (
+            obervarion[:max_length] + "... (truncated)"
+            if len(obervarion) > max_length
+            else obervarion
+        )
+
     @abstractmethod
     def generate_response(
         self,

dataherald/sql_generator/dataherald_finetuning_agent.py

@@ -228,7 +228,7 @@ async def _arun(
 class QuerySQLDataBaseTool(BaseSQLDatabaseTool, BaseTool):
     """Tool for querying a SQL database."""
 
-    name = "ExecuteQuery"
+    name = "SqlDbQuery"
     description = """
     Input: SQL query.
     Output: Result from the database or an error message if the query is incorrect.
@@ -591,6 +591,9 @@ def generate_response(
         response.sql = replace_unprocessable_characters(sql_query)
         response.tokens_used = cb.total_tokens
         response.completed_at = datetime.datetime.now()
+        response.intermediate_steps = self.construct_intermediate_steps(
+            result["intermediate_steps"], FINETUNING_AGENT_SUFFIX
+        )
         return self.create_sql_query_status(
             self.database,
             response.sql,

dataherald/sql_generator/dataherald_sqlagent.py

@@ -736,6 +736,13 @@ def generate_response(
         response.sql = replace_unprocessable_characters(sql_query)
         response.tokens_used = cb.total_tokens
         response.completed_at = datetime.datetime.now()
+        if number_of_samples > 0:
+            suffix = SUFFIX_WITH_FEW_SHOT_SAMPLES
+        else:
+            suffix = SUFFIX_WITHOUT_FEW_SHOT_SAMPLES
+        response.intermediate_steps = self.construct_intermediate_steps(
+            result["intermediate_steps"], suffix=suffix
+        )
         return self.create_sql_query_status(
             self.database,
             response.sql,

dataherald/types.py

@@ -177,13 +177,21 @@ class LLMConfig(BaseModel):
     api_base: str | None = None
 
 
+class IntermediateStep(BaseModel):
+    thought: str
+    action: str
+    action_input: str
+    observation: str
+
+
 class SQLGeneration(BaseModel):
     id: str | None = None
     prompt_id: str
     finetuning_id: str | None
     low_latency_mode: bool = False
     llm_config: LLMConfig | None
     evaluate: bool = False
+    intermediate_steps: list[IntermediateStep] | None
     sql: str | None
     status: str = "INVALID"
     completed_at: datetime | None

dataherald/utils/agent_prompts.py

@@ -121,7 +121,7 @@
 If SQL results has None or NULL values, handle them by adding a WHERE clause to filter them out.
 If SQL query doesn't follow the instructions or return incorrect results modify the SQL query to fit the instructions and fix the errors.
 Only make minor modifications to the SQL query, do not change the SQL query completely.
-You MUST always use the ExecuteQuery tool to make sure the SQL query is correct before returning it.
+You MUST always use the SqlDbQuery tool to make sure the SQL query is correct before returning it.
 
 ### Instructions from the database administrator:
 {admin_instructions}
@@ -134,7 +134,7 @@
 #
 Here is the plan you have to follow:
 1) Use the `GenerateSql` tool to generate a SQL query for the given question.
-2) Always Use the `ExecuteQuery` tool to execute the SQL query on the database to check if the results are correct.
+2) Always Use the `SqlDbQuery` tool to execute the SQL query on the database to check if the results are correct.
 #
 
 ### Instructions from the database administrator:

docs/api.create_prompt_sql_generation.rst

@@ -67,6 +67,13 @@ HTTP 201 code response
             "llm_name": "gpt-4-turbo-preview",
             "api_base": "string"
         },
+        "intermediate_steps": [
+            {
+                "action": "string",
+                "action_input": "string",
+                "observation": "string"
+            }
+        ],
         "sql": "string",
         "tokens_used": 0,
         "confidence_score": 0,
@@ -113,6 +120,15 @@ HTTP 201 code response
             "llm_name": "mistralai/Mixtral-8x7B-Instruct-v0.1",
             "api_base": "https://tt5h145hsc119q-8000.proxy.runpod.net/v1"
         },
+        intermediate_steps": [
+            {
+            "thought": "I should Collect examples of Question/SQL pairs to check if there is a similar question among the examples.\n",
+            "action": "FewshotExamplesRetriever",
+            "action_input": "5",
+            "observation": "samples ... "
+            },
+            ...
+        ],
         "sql": "SELECT metric_value \nFROM renthub_median_rent \nWHERE period_type = 'monthly' \nAND geo_type = 'city' \nAND location_name = 'Miami' \nAND property_type = 'All Residential' \nAND period_end = (SELECT DATE_TRUNC('MONTH', CURRENT_DATE()) - INTERVAL '1 day')\nLIMIT 10",
         "tokens_used": 18115,
         "confidence_score": 0.95,

docs/api.create_sql_generation.rst

@@ -61,6 +61,13 @@ HTTP 201 code response
             "llm_name": "gpt-4-turbo-preview",
             "api_base": "string"
         },
+        "intermediate_steps": [
+            {
+                "action": "string",
+                "action_input": "string",
+                "observation": "string"
+            }
+        ],
         "sql": "string",
         "tokens_used": 0,
         "confidence_score": 0,
@@ -102,6 +109,15 @@ HTTP 201 code response
         "llm_name": "mistralai/Mixtral-8x7B-Instruct-v0.1",
         "api_base": "https://tt5h145hsc119q-8000.proxy.runpod.net/v1"
     },
+    intermediate_steps": [
+            {
+            "thought": "I should Collect examples of Question/SQL pairs to check if there is a similar question among the examples.\n",
+            "action": "FewshotExamplesRetriever",
+            "action_input": "5",
+            "observation": "samples ... "
+            },
+            ...
+    ],
     "sql": "SELECT metric_value \nFROM renthub_median_rent \nWHERE period_type = 'monthly' \nAND geo_type = 'city' \nAND location_name = 'Miami' \nAND property_type = 'All Residential' \nAND period_end = (SELECT DATE_TRUNC('MONTH', CURRENT_DATE()) - INTERVAL '1 day')\nLIMIT 10",
     "tokens_used": 18115,
     "confidence_score": null,

docs/api.get_sql_generation.rst

@@ -28,6 +28,17 @@ HTTP 200 code response
     "finetuning_id": "string",
     "status": "string",
     "completed_at": "string",
+    "llm_config": {
+      "llm_name": "gpt-4-turbo-preview",
+      "api_base": "string"
+    },
+    "intermediate_steps": [
+      {
+        "action": "string",
+        "action_input": "string",
+        "observation": "string"
+      }
+    ],
     "sql": "string",
     "tokens_used": 0,
     "confidence_score": 0,
@@ -54,6 +65,19 @@ HTTP 200 code response
     "finetuning_id": null,
     "status": "VALID",
     "completed_at": "2024-01-04 21:11:27.235000+00:00",
+    "llm_config": {
+      "llm_name": "gpt-4-turbo-preview",
+      "api_base": null
+    },
+    "intermediate_steps": [
+      {
+        "thought": "I should Collect examples of Question/SQL pairs to check if there is a similar question among the examples.\n",
+        "action": "FewshotExamplesRetriever",
+        "action_input": "5",
+        "observation": "Found 5 examples of similar questions."
+        },
+        ...
+    ],
     "sql": "\nSELECT dh_zip_code, MAX(metric_value) as highest_rent -- Select the zip code and the maximum rent value\nFROM renthub_average_rent\nWHERE dh_county_name = 'Los Angeles' -- Filter for Los Angeles county\nAND period_start <= '2022-05-01' -- Filter for the period that starts on or before May 1st, 2022\nAND period_end >= '2022-05-31' -- Filter for the period that ends on or after May 31st, 2022\nGROUP BY dh_zip_code -- Group by zip code to aggregate rent values\nORDER BY highest_rent DESC -- Order by the highest rent in descending order\nLIMIT 1; -- Limit to the top result\n",
     "tokens_used": 12185,
     "confidence_score": null,

docs/api.list_sql_generations.rst

@@ -30,6 +30,17 @@ HTTP 200 code response
             "finetuning_id": "string",
             "status": "string",
             "completed_at": "string",
+            "llm_config": {
+                "llm_name": "gpt-4-turbo-preview",
+                "api_base": "string"
+            },
+            "intermediate_steps": [
+                {
+                    "action": "string",
+                    "action_input": "string",
+                    "observation": "string"
+                }
+            ],
             "sql": "string",
             "tokens_used": 0,
             "confidence_score": 0,
@@ -58,6 +69,19 @@ HTTP 200 code response
             "finetuning_id": null,
             "status": "VALID",
             "completed_at": "2024-01-03 18:54:55.091000+00:00",
+            "llm_config": {
+                "llm_name": "gpt-4-turbo-preview",
+                "api_base": null
+            },
+            "intermediate_steps": [
+                {
+                    "thought": "I should Collect examples of Question/SQL pairs to check if there is a similar question among the examples.\n",
+                    "action": "FewshotExamplesRetriever",
+                    "action_input": "5",
+                    "observation": "Found 5 examples of similar questions."
+                    },
+                    ...
+            ],
             "sql": "\nSELECT metric_value -- Rent price\nFROM renthub_median_rent\nWHERE geo_type='city' -- Focusing on city-level data\n  AND dh_state_name = 'California' -- State is California\n  AND dh_place_name = 'Los Angeles' -- City is Los Angeles\n  AND period_start = '2023-06-01' -- Most recent data available\nORDER BY metric_value DESC -- In case there are multiple entries, order by price descending\nLIMIT 1; -- Only need the top result\n",
             "tokens_used": 9491,
             "confidence_score": null,