Spaces:
Sleeping
Sleeping
File size: 23,538 Bytes
75d7142 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 |
# agent.py
import os
import json
import logging
from typing import Dict, Any, List, Optional
from langgraph.graph import StateGraph, END
# from langgraph.checkpoint.memory import MemorySaver # Keep commented unless needed
# Import the specific Tavily tool function
from tools import search_with_tavily
import prompts # Keep prompts for analysis, evaluation, synthesis
from state import AgentState # Keep state definition
# --- LLM and Logging Setup ---
import google.generativeai as genai
from dotenv import load_dotenv
load_dotenv()
API_KEY = os.getenv("GOOGLE_API_KEY")
if not API_KEY:
raise ValueError("GOOGLE_API_KEY not found in environment variables.")
genai.configure(api_key=API_KEY)
# Use Gemini 1.5 Flash - check model availability and naming conventions
# Consider error handling for model creation if needed
try:
llm = genai.GenerativeModel('gemini-1.5-flash')
except Exception as e:
logging.critical(f"Failed to initialize Gemini Model: {e}")
raise # Re-raise the exception to stop execution if LLM fails
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - [%(funcName)s] - %(message)s')
# --- Helper Function for LLM Calls ---
# agent.py
# ... (imports and setup) ...
# ... (logging) ...
# --- Helper Function for LLM Calls ---
# agent.py
# ... (imports and setup) ...
# --- Helper Function for LLM Calls ---
def call_llm(prompt: str) -> Optional[str]:
""" Helper function to call the Gemini LLM and handle potential errors. """
try:
logging.debug(f"Calling LLM. Prompt length: {len(prompt)}")
generation_config = genai.types.GenerationConfig(
max_output_tokens=4096,
temperature=0.7
)
# safety_settings = [...] # Optional
response = llm.generate_content(
prompt,
generation_config=generation_config,
# safety_settings=safety_settings
)
# Enhanced response checking
if not response.candidates:
if response.prompt_feedback and response.prompt_feedback.block_reason:
reason = response.prompt_feedback.block_reason
logging.error(f"LLM call blocked by API. Reason: {reason}")
return f"Error: LLM call blocked due to {reason}"
else:
logging.warning("LLM returned no candidates and no blocking reason.")
return None # Or return an empty string "" ?
# *** CORE FIX: Check the NAME of the finish reason ***
finish_reason_enum = response.candidates[0].finish_reason
if finish_reason_enum.name != 'STOP':
# Log details if it's not STOP
finish_reason_val = finish_reason_enum.value
logging.warning(f"LLM response finished with non-STOP reason: {finish_reason_enum.name} (Value: {finish_reason_val})")
# Check safety ratings if finish reason wasn't STOP
safety_reason = "Unknown"
# Basic check for safety ratings presence
if hasattr(response.candidates[0], 'safety_ratings') and response.candidates[0].safety_ratings:
for rating in response.candidates[0].safety_ratings:
# Check for a 'blocked' attribute, common in newer APIs
if hasattr(rating, 'blocked') and rating.blocked:
safety_reason = rating.category.name
break
# Add other safety check logic if needed based on API specifics
return f"Error: LLM response ended unexpectedly (Reason: {finish_reason_enum.name}, Safety Block Detected: {safety_reason})"
# --- If reason IS 'STOP', proceed to text extraction ---
# Access text via the 'parts' list
if response.candidates[0].content and response.candidates[0].content.parts:
if response.candidates[0].content.parts:
result = response.candidates[0].content.parts[0].text
logging.debug(f"LLM response received. Length: {len(result)}")
return result
else:
logging.warning("LLM response has content object but parts list is empty.")
return "" # Return empty string for valid empty responses
else:
logging.warning("LLM returned no content parts but finished normally (STOP reason).")
return "" # Return empty string for valid empty responses
except AttributeError as e:
logging.error(f"AttributeError processing LLM response: {e}. Response structure might have changed.", exc_info=True)
return f"Error: Failed to process LLM response structure - {e}"
except Exception as e:
logging.error(f"Error calling LLM: {e}", exc_info=True) # Log traceback
return f"Error: LLM API call failed - {e}"
# --- Rest of agent.py remains the same ---
# ... (copy the rest of your existing agent.py code here) ...
# --- Rest of agent.py remains the same ---
# ... (analyze_query_node, tavily_search_node, etc.) ...
# --- Helper Function for JSON Parsing ---
def clean_json_response(llm_output: str) -> Optional[dict]:
""" Attempts to parse JSON from LLM output, handling markdown code blocks. """
if not llm_output or llm_output.startswith("Error:"): # Don't try to parse error messages
return None
try:
# Find the start and end of the JSON block, handling potential ```json fences
json_start = llm_output.find('{')
json_end = llm_output.rfind('}')
if json_start != -1 and json_end != -1 and json_end > json_start:
json_str = llm_output[json_start:json_end + 1]
# Further clean potential markdown fences if they wrap the brackets
if json_str.strip().startswith("```json"):
json_str = json_str.strip()[7:]
if json_str.strip().endswith("```"):
json_str = json_str.strip()[:-3]
return json.loads(json_str.strip())
else:
logging.error(f"Could not find valid JSON object delimiters {{}} in LLM output: {llm_output}")
return None
except json.JSONDecodeError as e:
logging.error(f"Failed to decode JSON from LLM output snippet: {e}\nOutput was: {llm_output}")
return None # Failed to parse
# --- Agent Nodes (Adapted for Tavily) ---
def analyze_query_node(state: AgentState) -> Dict[str, Any]:
""" Analyzes the user query to plan the research. (Mostly unchanged) """
logging.info("Node: Analyzing Query")
query = state['original_query']
# Use the prompt from prompts.py as before
try:
prompt = prompts.QUERY_ANALYZER_PROMPT.format(query=query)
except KeyError as e:
logging.critical(f"KeyError during prompt formatting in analyze_query_node: {e}. Check prompts.py.")
# Cannot proceed without a valid prompt
return {"error_log": ["Critical prompt formatting error in analyze_query_node."]}
llm_response = call_llm(prompt)
parsed_analysis = clean_json_response(llm_response)
if parsed_analysis and isinstance(parsed_analysis.get('search_queries'), list):
logging.info(f"Query Analysis successful. Initial search queries: {parsed_analysis['search_queries']}")
# Initialize state fields needed for the Tavily flow
initial_updates = {
"query_analysis": parsed_analysis,
"search_queries": parsed_analysis['search_queries'],
"tavily_results": [], # Store results from Tavily
"accumulated_report_notes": [], # Store formatted Tavily result content
"error_log": [],
"current_iteration": 0
}
current_state = state.copy()
current_state.update(initial_updates)
return current_state # Return the whole updated state dictionary
else:
logging.error(f"Failed to parse query analysis from LLM. Raw Response: {llm_response}")
error_msg = f"Failed to parse LLM response for query analysis. Raw response: {llm_response}"
error_log = state.get("error_log", []) + [error_msg]
# Return only the fields to update, LangGraph merges them
return {"error_log": error_log, "search_queries": []}
def tavily_search_node(state: AgentState) -> Dict[str, Any]:
""" Performs search using Tavily API. """
logging.info("Node: Tavily Search")
# Use .get() with defaults for robustness
search_queries = state.get('search_queries', [])
tavily_results_so_far = state.get('tavily_results', [])
accumulated_notes = state.get('accumulated_report_notes', [])
errors_so_far = state.get('error_log', [])
if not search_queries:
logging.warning("No search queries available for Tavily. Skipping search node.")
return {} # No change if no queries
# Use the first query in the list
query = search_queries[0]
remaining_queries = search_queries[1:] # Prepare for update later
# Call Tavily - use include_answer=True to get a potential synthesized answer
tavily_response = search_with_tavily(
query=query,
search_depth="basic", # Start with basic, consider "advanced" later if needed
max_results=5,
include_answer=True # Request Tavily's synthesized answer
)
# Initialize updates, default to no change
current_errors = errors_so_far
new_results = tavily_results_so_far
new_notes = accumulated_notes
# Process response
if tavily_response and "error" not in tavily_response:
# Get results safely using .get()
results_list = tavily_response.get('results', [])
tavily_answer = tavily_response.get('answer') # Can be None
# Append raw results (optional, good for debugging)
new_results.extend(results_list)
# Add Tavily's answer to notes if present and not empty
if tavily_answer:
note = f"Tavily Answer (for query: '{query}'):\n{tavily_answer}\n---\n"
new_notes.append(note)
logging.info("Added Tavily's synthesized answer to notes.")
# Add summaries from individual results to notes
if results_list:
for result in results_list:
# Safely get attributes from each result dictionary
url = result.get('url', 'N/A')
title = result.get('title', 'No Title')
content_summary = result.get('content', 'No Summary Provided')
note = f"Source: {url}\nTitle: {title}\nContent Summary: {content_summary}\n---\n"
new_notes.append(note)
logging.info(f"Added {len(results_list)} result summaries to notes.")
else:
logging.info("Tavily returned no individual results for this query.")
else:
# Log Tavily API errors
error_msg = tavily_response.get("error", f"Unknown error during Tavily search for '{query}'") if isinstance(tavily_response, dict) else f"Invalid Tavily response format for '{query}'"
logging.error(error_msg)
current_errors.append(error_msg)
# Return the dictionary of fields to update
return {
"tavily_results": new_results,
"accumulated_report_notes": new_notes,
"search_queries": remaining_queries, # Update the list
"error_log": current_errors
}
# Removed: filter_select_urls_node, scrape_websites_node, analyze_content_node
def evaluate_progress_node(state: AgentState) -> Dict[str, Any]:
""" Evaluates progress based on Tavily results and decides next step. """
logging.info("Node: Evaluate Progress (Tavily Flow)")
query = state['original_query']
analysis = state.get('query_analysis', {}) # Get safely
# Use notes accumulated from Tavily's answers and summaries
notes = "\n".join(state.get('accumulated_report_notes', ["No information gathered yet."]))
current_iter = state.get('current_iteration', 0)
max_iter = state['max_iterations']
# Prepare analysis JSON safely
try:
analysis_json = json.dumps(analysis, indent=2) if analysis else "{}"
except TypeError as e:
logging.error(f"Could not serialize query analysis to JSON: {e}")
analysis_json = "{}" # Fallback
# Prompt LLM to evaluate based on Tavily results in notes
try:
prompt = prompts.EVALUATOR_PROMPT.format(
query=query,
analysis=analysis_json,
notes=notes,
iteration=current_iter,
max_iterations=max_iter
)
except KeyError as e:
logging.critical(f"KeyError during prompt formatting in evaluate_progress_node: {e}. Check prompts.py.")
# Need to make a decision even if prompt fails
error_log = state.get("error_log", []) + [f"Critical prompt formatting error in evaluate_progress_node: {e}"]
return {"error_log": error_log, "_decision": "stop", "current_iteration": current_iter + 1}
llm_response = call_llm(prompt)
parsed_eval = clean_json_response(llm_response)
decision = "stop" # Default to stopping
next_queries = []
if parsed_eval and isinstance(parsed_eval.get('decision'), str): # Basic validation
decision = parsed_eval['decision'].lower()
if decision == 'continue':
next_queries_raw = parsed_eval.get('next_search_queries', [])
# Ensure next_queries is a list of strings
next_queries = [q for q in next_queries_raw if isinstance(q, str)] if isinstance(next_queries_raw, list) else []
if not next_queries:
logging.warning("Evaluator decided to continue but provided no valid new queries. Will stop.")
decision = 'stop'
else:
logging.info(f"Evaluator decided to continue. New queries for Tavily: {next_queries}")
elif decision == 'synthesize':
logging.info("Evaluator decided to synthesize.")
else: # stop or invalid decision string
if decision != 'stop':
logging.warning(f"Invalid decision '{decision}' received from evaluator LLM. Defaulting to stop.")
decision = 'stop'
logging.info(f"Evaluator decided to stop. Reason: {parsed_eval.get('assessment', 'N/A')}")
else:
logging.error(f"Failed to parse evaluation response or get valid decision from LLM. Stopping. Raw: {llm_response}")
error_log = state.get("error_log", []) + [f"Failed to parse evaluation response or get valid decision. Raw: {llm_response}"]
return {"error_log": error_log, "_decision": "stop", "current_iteration": current_iter + 1}
# Update state for next loop or final step
updates = {"current_iteration": current_iter + 1, "_decision": decision}
if decision == 'continue':
# Prepend new queries to the list to be processed next
current_queries = state.get('search_queries', [])
updates["search_queries"] = next_queries + current_queries # Combine lists
return updates
def synthesize_report_node(state: AgentState) -> Dict[str, Any]:
""" Generates the final research report based on Tavily results. """
logging.info("Node: Synthesize Final Report (Tavily Flow)")
query = state['original_query']
analysis = state.get('query_analysis', {}) # Get safely
# Notes now contain Tavily answers/summaries
notes = "\n".join(state.get('accumulated_report_notes', ["No information gathered."]))
errors = "\n".join(state.get('error_log', ["None"])) # Get safely
# Prepare analysis JSON safely
try:
analysis_json = json.dumps(analysis, indent=2) if analysis else "{}"
except TypeError as e:
logging.error(f"Could not serialize query analysis to JSON for synthesis: {e}")
analysis_json = "{}" # Fallback
# Use the existing SYNTHESIS_PROMPT - it takes notes and should work
try:
prompt = prompts.SYNTHESIS_PROMPT.format(
query=query, analysis=analysis_json, notes=notes, errors=errors
)
except KeyError as e:
logging.critical(f"KeyError during prompt formatting in synthesize_report_node: {e}. Check prompts.py.")
fallback = f"Critical prompt formatting error during synthesis. Review notes:\n{notes}\nErrors:\n{errors}"
error_log = state.get("error_log", []) + [f"Critical prompt formatting error in synthesize_report_node: {e}"]
return {"final_report": fallback, "error_log": error_log}
final_report = call_llm(prompt)
# Check if LLM call itself returned an error string
if final_report and final_report.startswith("Error:"):
logging.error(f"Failed to generate final report. LLM Error: {final_report}")
fallback = f"Failed to synthesize report due to LLM error ({final_report}). Please review accumulated notes:\n{notes}\nErrors:\n{errors}"
error_log = state.get("error_log", []) + [f"Synthesis failed. LLM Error: {final_report}"]
return {"final_report": fallback, "error_log": error_log}
elif not final_report: # Handle None or empty string case
logging.error("Failed to generate final report. LLM returned empty content.")
fallback = f"Failed to synthesize report (LLM returned empty content). Please review accumulated notes:\n{notes}\nErrors:\n{errors}"
error_log = state.get("error_log", []) + ["Synthesis failed: LLM returned empty content."]
return {"final_report": fallback, "error_log": error_log}
else:
# Success case
logging.info("Successfully generated final report.")
return {"final_report": final_report}
# --- Conditional Edge Logic ---
def route_after_evaluation(state: AgentState) -> str:
""" Determines the next node based on the evaluation decision. """
# Use .get for safety
decision = state.get("_decision")
current_iter = state.get("current_iteration", 0)
max_iter = state.get("max_iterations", 3) # Default if not set
search_queries_left = state.get("search_queries", [])
logging.debug(f"Routing: Decision='{decision}', Iteration={current_iter}/{max_iter}, Queries Left={len(search_queries_left)}")
# Check if max iterations reached OR if decision is continue BUT no queries left
if current_iter >= max_iter:
logging.warning(f"Max iterations ({max_iter}) reached. Forcing synthesis.")
return "synthesize" # Force synthesis
elif decision == "continue" and not search_queries_left:
logging.warning("Decision was 'continue' but no search queries remain. Forcing synthesis.")
return "synthesize" # Force synthesis
if decision == "continue":
# Route back to Tavily search node
return "continue_search"
elif decision == "synthesize":
return "synthesize"
else: # stop or error or invalid decision string
logging.info(f"Routing to synthesize based on decision '{decision}' or error.")
return "synthesize"
# --- Build the Graph (Adapted for Tavily Flow) ---
def create_graph() -> StateGraph:
""" Creates and configures the LangGraph agent with Tavily. """
workflow = StateGraph(AgentState)
# Add nodes for the new flow
workflow.add_node("analyze_query", analyze_query_node)
workflow.add_node("tavily_search", tavily_search_node)
# Removed: filter_select_urls, scrape_websites, analyze_content
workflow.add_node("evaluate_progress", evaluate_progress_node)
workflow.add_node("synthesize_report", synthesize_report_node)
# Define edges for the new flow
workflow.set_entry_point("analyze_query")
workflow.add_edge("analyze_query", "tavily_search")
# Removed edges related to scraping/filtering/content analysis
# Edge from search directly to evaluation
workflow.add_edge("tavily_search", "evaluate_progress")
# Conditional edge from evaluation
workflow.add_conditional_edges(
"evaluate_progress",
route_after_evaluation,
{
"continue_search": "tavily_search", # Loop back to Tavily search
"synthesize": "synthesize_report", # Move to synthesis
}
)
workflow.add_edge("synthesize_report", END)
# Compile the graph - consider adding checkpointing for resilience
# memory = MemorySaver()
# app = workflow.compile(checkpointer=memory)
app = workflow.compile()
logging.info("LangGraph agent graph compiled for Tavily flow.")
return app
# --- Main Agent Class (Remains the same structure) ---
class ResearchAgent:
def __init__(self, max_iterations=3):
self.app = create_graph()
self.max_iterations = max_iterations
logging.info(f"Research Agent initialized with max_iterations={max_iterations} (Tavily Flow).")
def run(self, query: str) -> Dict[str, Any]:
if not query or not query.strip(): # Check for empty/whitespace query
logging.error("Query cannot be empty.")
return {"error": "Query cannot be empty.", "final_report": "Error: Query cannot be empty."}
initial_state = AgentState(
original_query=query,
max_iterations=self.max_iterations,
# Initialize fields potentially used in the graph
query_analysis=None,
search_queries=[],
tavily_results=[],
accumulated_report_notes=[],
final_report=None,
error_log=[],
current_iteration=0,
# Include other keys defined in AgentState with default values
# even if not directly used in the main Tavily flow,
# to prevent potential key errors if accessed unexpectedly.
search_results=[],
urls_to_scrape=[],
scraped_data={},
analyzed_data={},
visited_urls=set()
)
logging.info(f"Starting research for query: '{query}' (Tavily Flow)")
# Increase recursion limit for potential loops
config = {"recursion_limit": 50}
final_state = {} # Initialize final_state
try:
final_state = self.app.invoke(initial_state, config=config)
logging.info("Research process finished (Tavily Flow).")
except Exception as e:
logging.critical(f"LangGraph invocation failed: {e}", exc_info=True)
# Populate final_state with error information
final_state = initial_state # Start with initial state
final_state['error_log'] = final_state.get('error_log', []) + [f"CRITICAL: Agent execution failed: {e}"]
final_state['final_report'] = f"CRITICAL ERROR: Agent execution failed. Check logs. Error: {e}"
# Clean up temporary keys before returning, using .pop with default None
final_state.pop('_decision', None)
# Ensure essential keys exist in the returned state, even if run failed early
final_state.setdefault('final_report', "Processing failed before report generation.")
final_state.setdefault('error_log', [])
return final_state |