Update app.py

app.py

@@ -7,6 +7,9 @@ from smolagents import OpenAIServerModel
 from smolagents import CodeAgent, Tool, tool
 from smolagents import DuckDuckGoSearchTool, VisitWebpageTool
 from smolagents import PythonInterpreterTool # Import the built-in Python Interpreter Tool
+from smolagents.agents.base import LogLevel # Import LogLevel for verbosity
+import time # Import time for sleep
+from requests.exceptions import HTTPError # Import HTTPError for specific exception handling
 
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
@@ -19,8 +22,8 @@ class GaiaFileTool(Tool):
     """
     name = "download_gaia_file"
     description = "Downloads a file associated with a given GAIA task ID and returns its content. It takes 'task_id' as input and returns the file content as a string. Use this when a question refers to an external file."
-    inputs = {"task_id": {"type": "string", "description": "The task ID for which to download the file (e.g., '2345')."}} # Corrected type to "string"
-    output_type = "string" # Corrected type to "string"
+    inputs = {"task_id": {"type": "string", "description": "The task ID for which to download the file (e.g., '2345')."}}
+    output_type = "string"
 
     def __init__(self, api_base_url=DEFAULT_API_URL):
         super().__init__()
@@ -43,8 +46,6 @@ class GaiaFileTool(Tool):
             print(f"Error downloading file for task_id {task_id}: {e}")
             return f"Error downloading file: {e}"
 
-# Removed the custom python_repl function as we are using the built-in tool
-
 # --- Custom GAIA Agent Definition ---
 class GaiaAgent(CodeAgent):
     """
@@ -74,8 +75,6 @@ class GaiaAgent(CodeAgent):
         visit_webpage_tool = VisitWebpageTool()
 
         # Initialize the built-in Python Interpreter Tool
-        # By default, PythonInterpreterTool uses a local executor, which is generally safe
-        # for controlled environments like Hugging Face Spaces.
         python_interpreter_tool = PythonInterpreterTool()
 
         # Define the tools available to the agent
@@ -85,38 +84,95 @@ class GaiaAgent(CodeAgent):
             duckduckgo_search_tool,
             visit_webpage_tool
         ]
-        super().__init__(model=self.llm_model, tools=agent_tools)
+        super().__init__(model=self.llm_model, tools=agent_tools, verbosity_level=LogLevel.DEBUG)
         print("GaiaAgent initialized successfully with Gemini Flash and built-in tools.")
 
     def __call__(self, question: str) -> str:
         """
         The main method for the agent to process a question and return an answer.
         This will involve the agent's internal reasoning, tool use, and planning.
+        Includes retry logic for LLM calls to handle rate limits.
         """
         print(f"\n--- Agent received question (first 100 chars): {question[:100]}...")
 
-        try:
-            prompt = (
-                f"You are an AI agent designed to solve GAIA benchmark questions. "
-                f"Your goal is to provide the exact answer as a string, without any additional text, "
-                f"explanation, or the phrase 'FINAL ANSWER:'. "
-                f"Break down the problem, use the available tools (python_interpreter, download_gaia_file, "
-                f"duckduckgo_search_tool, visit_webpage_tool) as needed, and think step-by-step. "
-                f"Use 'python_interpreter' for any calculations or code execution. "
-                f"Use 'duckduckgo_search_tool' to find information on the web. "
-                f"Use 'visit_webpage_tool' to read the content of a specific URL. "
-                f"When you have the final answer, output ONLY the answer string.\n\n"
-                f"Question: {question}"
-            )
-            
-            result = self.run(prompt)
-
-            final_answer = self._extract_exact_answer(result)
-            print(f"--- Agent returning final answer (first 100 chars): {final_answer[:100]}...")
-            return final_answer
-        except Exception as e:
-            print(f"--- Error during agent execution: {e}")
-            return "Agent encountered an error and could not provide an answer."
+        prompt = (
+            f"You are an AI agent designed to solve GAIA benchmark questions. "
+            f"Your goal is to provide the exact answer as a string, without any additional text, "
+            f"explanation, or the phrase 'FINAL ANSWER:'. "
+            f"Break down the problem, use the available tools (python_interpreter, download_gaia_file, "
+            f"duckduckgo_search_tool, visit_webpage_tool) as needed, and think step-by-step. "
+            f"When using web search or webpage visit tools, be highly efficient. "
+            f"Formulate comprehensive search queries to get as much relevant information as possible in one go. "
+            f"Only visit a webpage if absolutely necessary and when you expect it to contain the direct answer or crucial data. "
+            f"Avoid redundant searches or visiting multiple pages for the same piece of information. "
+            f"Use 'python_interpreter' for any calculations or code execution. "
+            f"Use 'duckduckgo_search_tool' to find information on the web. "
+            f"Use 'visit_webpage_tool' to read the content of a specific URL. "
+            f"When you have the final answer, output ONLY the answer string.\n\n"
+            f"Question: {question}"
+        )
+        
+        print(f"Agent running with prompt (first 200 chars): {prompt[:200]}...")
+
+        max_retries = 5 
+        initial_retry_delay = 30 
+        retry_delay = initial_retry_delay
+        result = None
+
+        for attempt in range(max_retries):
+            try:
+                result = self.run(prompt)
+                print(f"Agent raw output from self.run():\n{result}") # Log the raw output
+                break # Break loop if successful
+            except HTTPError as e:
+                if e.response.status_code == 429:
+                    error_details = ""
+                    try:
+                        # Attempt to parse more specific error details from the response
+                        error_json = e.response.json()
+                        if 'error' in error_json and 'details' in error_json['error']:
+                            for detail in error_json['error']['details']:
+                                if detail.get('@type') == 'type.googleapis.com/google.rpc.QuotaFailure':
+                                    quota_metric = detail.get('quotaMetric', 'N/A')
+                                    quota_id = detail.get('quotaId', 'N/A')
+                                    quota_value = detail.get('quotaValue', 'N/A')
+                                    error_details = f"Quota Metric: {quota_metric}, Quota ID: {quota_id}, Value: {quota_value}. "
+                                    break # Found relevant detail
+                    except Exception as parse_error:
+                        print(f"Could not parse detailed error from 429 response: {parse_error}")
+                        error_details = "Check Google Cloud Console for details. "
+
+                    error_message = (
+                        f"Gemini API Rate limit hit (429) on attempt {attempt + 1}/{max_retries}. "
+                        f"{error_details}"
+                        f"Retrying in {retry_delay} seconds... "
+                        f"This could be due to the 15 RPM or 200 RPD free tier limits. "
+                        f"If this persists, your daily quota might be exhausted."
+                    )
+                    print(error_message)
+                    time.sleep(retry_delay)
+                    retry_delay *= 2 # Exponential backoff
+                else:
+                    # Re-raise other HTTP errors
+                    raise
+            except Exception as e:
+                # Log the full traceback for better debugging
+                import traceback
+                print(f"--- Error during agent execution on attempt {attempt + 1}/{max_retries}: {e}")
+                traceback.print_exc()
+                if attempt < max_retries - 1:
+                    print(f"Retrying in {retry_delay} seconds...")
+                    time.sleep(retry_delay)
+                    retry_delay *= 2
+                else:
+                    return "Agent encountered an error and could not provide an answer after multiple retries." # Final failure after retries
+
+        if result is None:
+            return "Agent failed after multiple retries due to an unknown error or persistent rate limits."
+
+        final_answer = self._extract_exact_answer(result)
+        print(f"--- Agent returning final answer (first 100 chars): {final_answer[:100]}...")
+        return final_answer
 
     def _extract_exact_answer(self, raw_output: str) -> str:
         """
@@ -124,6 +180,8 @@ class GaiaAgent(CodeAgent):
         Ensures no "FINAL ANSWER" text is included and handles any
         extraneous formatting. This function is crucial for GAIA's exact match scoring.
         """
+        print(f"Attempting to extract exact answer from raw output (first 200 chars):\n{raw_output[:200]}...")
+
         cleaned_output = raw_output.replace("FINAL ANSWER:", "").strip()
         cleaned_output = cleaned_output.replace("Answer:", "").strip()
         cleaned_output = cleaned_output.replace("The answer is:", "").strip()
@@ -144,6 +202,7 @@ class GaiaAgent(CodeAgent):
         if cleaned_output.startswith("'") and cleaned_output.endswith("'"):
             cleaned_output = cleaned_output[1:-1]
 
+        print(f"Extracted and cleaned answer: {cleaned_output[:100]}...")
         return cleaned_output.strip()
 
 
@@ -170,6 +229,9 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
     try:
         agent = GaiaAgent()
     except Exception as e:
+        print(f"Error during agent initialization in run_and_submit_all: {e}")
+        import traceback
+        traceback.print_exc()
         return f"Error initializing agent: {e}", None
 
     try:
@@ -179,6 +241,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         questions = questions_response.json()
         print(f"Fetched {len(questions)} questions.")
     except requests.exceptions.RequestException as e:
+        print(f"Error fetching questions: {e}")
         return f"Error fetching questions: {e}", None
 
     all_answers = []