Changed to mini models of openai to make it faster

agents/llama_index_agent.py

@@ -83,11 +83,14 @@ class GaiaAgent(ReActAgent):
         if system_prompt is None:
             system_prompt = self._get_default_system_prompt()
 
+        # CRITICAL: Explicitly define which agents this agent can hand off to
+        # This needs to be defined before calling super().__init__
         can_handoff_to = [
             "writer_agent",
+            "review_agent"
             ]
             
-        # Initialize the parent ReActAgent
+        # Initialize the parent ReActAgent with explicit handoff configuration
         super().__init__(
             name=name,
             description=description,
@@ -95,6 +98,7 @@ class GaiaAgent(ReActAgent):
             system_prompt=system_prompt,
             tools=tools,
             can_handoff_to=can_handoff_to,
+            verbose=True,  # Enable verbose mode for better debugging
             **kwargs
         )
         
@@ -114,7 +118,7 @@ class GaiaAgent(ReActAgent):
             return Anthropic(
                 model=model_name,
                 api_key=api_key or os.getenv("ANTHROPIC_API_KEY"),
-                temperature=1.0 if "3-7" in model_name else 0.5,
+                temperature=0.7 if "3-7" in model_name else 0.5,  # Slightly reduced temperature
                 thinking_dict={"type": "enabled", "budget_tokens": 2048} if "3-7" in model_name else None,
                 max_tokens=2048*4
             )
@@ -129,6 +133,14 @@ class GaiaAgent(ReActAgent):
         return """
         You are the lead coordinator for a team of specialized AI agents tackling the GAIA benchmark. Your job is to analyze questions and generate detailed analysis, which you'll pass to a specialized formatting agent for final answer preparation.
 
+        ## CRITICAL WORKFLOW INSTRUCTIONS
+        YOU MUST FOLLOW THIS EXACT WORKFLOW:
+        1. Analyze the question thoroughly
+        2. Use tools if needed for research
+        3. Formulate your answer
+        4. ALWAYS use the handoff tool to delegate to writer_agent
+        5. NEVER provide a direct answer with "Answer:" - this breaks the workflow
+
         ## QUESTION ANALYSIS PROCESS
         1. First, carefully read and parse the entire question
         2. Identify the EXACT output format required (single word, name, number, comma-separated list, etc.)
@@ -185,25 +197,17 @@ class GaiaAgent(ReActAgent):
         4. For data analysis tasks, ensure you've properly processed the CSV data and extracted the requested information
         5. When calculations or statistics are needed, perform them accurately and document your methodology
 
-        ## DELEGATION TO WRITER AGENT
-        After completing your analysis, ALWAYS delegate the final answer preparation to the writer_agent with:
-        - query: The original question
-        - research_notes: Your complete analysis, all relevant facts, and what you believe is the correct answer
-        - answer_format: EXPLICIT instructions on exactly how the answer should be formatted (single word, comma-separated list, etc.)
-
-        Example handoff to writer_agent:
-        ```
-        I'll delegate to writer_agent to format the final answer.
-        
-        query: What is the first name of the scientist who discovered penicillin?
-        research_notes: After researching, I found that Sir Alexander Fleming discovered penicillin in 1928. The full answer is "Alexander Fleming" but the question only asks for the first name, which is "Alexander".
-        ```
+        ## REQUIRED HANDOFF TO WRITER AGENT
+        After completing your analysis, you MUST ALWAYS use the handoff tool with EXACTLY this format:
 
-        IMPORTANT: NEVER provide the final answer directly to the user. ALWAYS hand off to the writer_agent for proper formatting.
+        Thought: I have completed my analysis and need to delegate to the writer agent for proper formatting.
+        Action: handoff
+        Action Input: {"to_agent": "writer_agent", "reason": "Need to format the final answer according to requirements", "query": "<original question>", "research_notes": "<my complete analysis and what I believe is the correct answer>", "format_requirements": "<explicit formatting instructions>"}
 
-        IMPORTANT: The final answer will be provided by the `review_agent`. Publish it as you receive it.
+        CRITICAL: Even if the answer seems obvious or simple, you MUST use the handoff tool with the EXACT format above. NEVER respond with "Answer: ..." as this breaks the workflow.
         """
 
+
 def create_writer_agent(model_config: Dict[str, Any]) -> ReActAgent:
     """
     Create a writer agent that formats final answers based on research notes.
@@ -241,46 +245,49 @@ def create_writer_agent(model_config: Dict[str, Any]) -> ReActAgent:
     else:
         raise ValueError(f"Unsupported model provider for writer agent: {model_provider}")
     
-    # Create and return the writer agent
+    # Create and return the writer agent with updated system prompt
     return ReActAgent(
         name="writer_agent",
         description="Formats the final answer based on research notes for GAIA benchmark questions",
         system_prompt="""
         You are a specialized formatting agent for the GAIA benchmark. Your job is to take the research from the main agent and format the answer according to the benchmark requirements.
 
+        ## CRITICAL WORKFLOW INSTRUCTIONS
+        YOU MUST FOLLOW THIS EXACT WORKFLOW:
+        1. Review the main agent's research and format requirements
+        2. Format the answer according to specifications
+        3. ALWAYS use the handoff tool to delegate to review_agent
+        4. NEVER provide a direct answer with "Answer:" - this breaks the workflow
+
         ## YOUR ROLE
         You will receive:
         - query: The original question
         - research_notes: The main agent's complete analysis and reasoning
+        - format_requirements: How the answer should be formatted
 
         ## FORMATTING RULES
-        1. Format the answer according to the instructions in the `query` received
+        1. Format the answer according to the instructions in the `query` and `format_requirements`
         2. Your answers will be always as minimal as necessary to answer the question
-        2. Try to remove unnecessary characters, spaces, or wording
-        3. If asked for a name, provide **ONLY** the name
-        4. If asked for a number, provide the **ONLY** number
-        5. If asked for a list, format it exactly as specified
+        3. Remove unnecessary characters, spaces, or wording
+        4. If asked for a name, provide **ONLY** the name
+        5. If asked for a number, provide the **ONLY** number
+        6. If asked for a list, format it exactly as specified
 
-        ## DELEGATION TO REVIEW AGENT
-        After formatting your answer, ALWAYS delegate to the review_agent with:
-        - query: The original question
-        - formatted_answer: Your formatted answer
+        ## REQUIRED HANDOFF TO REVIEW AGENT
+        After formatting your answer, you MUST ALWAYS use the handoff tool with EXACTLY this format:
 
-        Example handoff to review_agent:
-        ```
-        I'll delegate to review_agent for final review.
-        
-        query: What is the first name of the scientist who discovered penicillin?
-        formatted_answer: Alexander
-        format_requirements: Return ONLY the first name, with no additional text.
-        ```
+        Thought: I have formatted the answer and need to delegate to the review agent for final verification.
+        Action: handoff
+        Action Input: {"to_agent": "review_agent", "reason": "Need final verification and cleanup", "query": "<original question>", "formatted_answer": "<my formatted answer>", "format_requirements": "<explicit formatting requirements>"}
 
-        IMPORTANT: ALWAYS hand off to the "review_agent" for final verification and cleanup.
+        CRITICAL: Even if the answer seems perfectly formatted already, you MUST use the handoff tool with the EXACT format above. NEVER respond with "Answer: ..." as this breaks the workflow.
         """,
         llm=llm,
-        can_handoff_to=["review_agent"]
+        can_handoff_to=["review_agent"],
+        verbose=True  # Enable verbose mode for better debugging
     )
 
+
 def create_review_agent(model_config: Dict[str, Any]) -> ReActAgent:
     """
     Create a review agent that ensures the final answer follows exact formatting requirements.
@@ -329,6 +336,7 @@ def create_review_agent(model_config: Dict[str, Any]) -> ReActAgent:
         You will receive:
         - query: The original question
         - formatted_answer: The answer formatted by the writer agent
+        - format_requirements: How the answer should be formatted
 
         ## CRITICAL RULES
         1. Your ENTIRE response must be ONLY the final answer - NOTHING ELSE
@@ -347,6 +355,9 @@ def create_review_agent(model_config: Dict[str, Any]) -> ReActAgent:
            - Case sensitivity (e.g., "PARIS" vs "Paris")
            - List formatting (e.g., comma-separated vs numbered)
 
+        ## OUTPUT FORMAT
+        Your entire response should be ONLY the final, formatted answer. Do not include any additional text, explanation, or reasoning.
+
         ## OUTPUT EXAMPLES
         - Input: "The answer is Alexander."
           Output: Alexander
@@ -361,10 +372,10 @@ def create_review_agent(model_config: Dict[str, Any]) -> ReActAgent:
         - Input: "She published studio albums "Album 1", "Album 2", "Album 3", so in total 3."
           Output: 3
         - Input: Misa Criolla (2000), Acústico (2002), Corazón Libre (2005), Cantora 1 (2009), and Cantora 2 (2009)
-          Output 5
+          Output: 5
 
         REMEMBER: Your ENTIRE response should be just the bare answer with NOTHING else.
         """,
         llm=llm,
-        can_handoff_to=["jefe"]
-    )
+        can_handoff_to=[]  # Review agent is the final step, so it doesn't hand off to anyone
+    )

app.py

@@ -32,8 +32,8 @@ OPENAI = {
 class BasicAgent:
     def __init__(
             self,
-            model_provider="anthropic",
-            model_name="claude-3-7-sonnet-latest",
+            model_provider="openai",
+            model_name="o4-mini",
             api_key=None,
             use_separate_writer_model=True,
             writer_model_provider="openai",
@@ -126,7 +126,7 @@ class BasicAgent:
         
         print(f"Agent received question (first 50 chars): {question_text[:50]}...")
         
-        # Download audio file if present
+        # Download file if present
         local_file_path = None
         if file_name and task_id:
             try:
@@ -144,30 +144,58 @@ class BasicAgent:
                 "analysis_notes": "",
                 "format_requirements": "",
                 "next_agent": "",
-                "final_answer": ""
+                "final_answer": "",
+                "workflow_state": "initial_analysis",  # Track workflow state
+                "require_handoff": True,              # Flag that handoff is required
             }
             
-            # MODIFY THIS PART - Instead of just passing the question text,
-            # create a more detailed input that includes the audio file path information
-            enhanced_input = f"""Task ID: {task_id}
-                Question: {question_text}
-
+            # Create a more detailed input with workflow instructions
+            enhanced_input = f"""
+            WORKFLOW INSTRUCTIONS:
+            1. You (jefe) MUST analyze this question and find the answer
+            2. After analysis, you MUST use the handoff tool to delegate to writer_agent
+            3. NEVER provide a direct answer - always delegate using the handoff tool
+            
+            Task ID: {task_id}
+            Question: {question_text}
             """
-            # Add audio file information if available
-            if local_file_path:
-                enhanced_input += f"Audio File Path: {local_file_path}\n\nPlease analyze this question. If it involves an audio file, use the transcribe_audio tool with the provided path."
             
-            # Use the workflow to process the question with enhanced input
-            workflow_response = await self.agent_workflow.run(
-                enhanced_input,
-                initial_state=initial_state
-            )
-            return workflow_response
+            # Add file information if available
+            if local_file_path:
+                enhanced_input += f"\nFile Path: {local_file_path}\n\nPlease analyze this question. If it involves an audio file, use the transcribe_audio tool with the provided path."
             
+            # Monitor the workflow execution
+            print("Starting workflow execution...")
+            try:
+                workflow_response = await self.agent_workflow.run(
+                    enhanced_input,
+                    initial_state=initial_state
+                )
+                
+                # Extract the final answer from the last response
+                if hasattr(workflow_response.response, 'blocks') and workflow_response.response.blocks:
+                    final_answer = workflow_response.response.blocks[-1].text
+                    print(f"Workflow completed. Final answer extracted: {final_answer}")
+                    return final_answer
+                else:
+                    print("Warning: Could not extract final answer from workflow response blocks")
+                    # Try to extract from the response content
+                    final_answer = str(workflow_response.response)
+                    return final_answer
+                    
+            except Exception as e:
+                print(f"Error in workflow execution: {e}")
+                import traceback
+                traceback.print_exc()
+                return f"Error: {str(e)}"
+        
         response = asyncio.run(agentic_main())
         
-        # Extract the final answer from the writer agent's response
-        final_answer = response.response.blocks[-1].text
+        # Extract the final answer and remove any "Answer:" prefix
+        final_answer = response.response.blocks[-1].text if hasattr(response, 'response') and hasattr(response.response, 'blocks') else str(response)
+        if isinstance(final_answer, str) and final_answer.startswith("Answer:"):
+            final_answer = final_answer.replace("Answer:", "").strip()
+        
         print(f"Agent returning final answer: {final_answer}")
         return final_answer