Update app.py

app.py

@@ -4,8 +4,7 @@ import gradio as gr
 import requests
 import pandas as pd
 
-from smolagents import Tool, CodeAgent, HfApiModel
-
+from smolagents import InferenceClientModel, ToolCallingAgent
 from audio_transcriber import AudioTranscriptionTool
 from image_analyzer import ImageAnalysisTool
 from wikipedia_searcher import WikipediaSearcher
@@ -13,46 +12,68 @@ from wikipedia_searcher import WikipediaSearcher
 # GAIA scoring endpoint
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
-# Static system prompt for GAIA
-SYSTEM_PROMPT = """You are an agent solving the GAIA benchmark and you are required to provide exact answers.
+# Define the GaiaAgent class with embedded prompt in __call__
+class GaiaAgent:
+    def __init__(self):
+        print("Gaia Agent Initialized")
+
+        self.model = InferenceClientModel(
+            model_id="cognitivecomputations/dolphin-2.6-mixtral-8x7b",
+            token=os.getenv("HUGGINGFACEHUB_API_TOKEN", "").strip()
+        )
+
+        self.tools = [
+            AudioTranscriptionTool(),
+            ImageAnalysisTool(),
+            WikipediaSearcher()
+        ]
+
+        self.agent = ToolCallingAgent(
+            tools=self.tools,
+            model=self.model
+        )
+
+    def __call__(self, question: str) -> str:
+        print(f"Agent received question (first 50 chars): {question[:50]}...")
+
+        prompt = f"""You are an agent solving the GAIA benchmark and you are required to provide exact answers.
 Rules to follow:
 1. Return only the exact requested answer: no explanation and no reasoning.
 2. For yes/no questions, return exactly \"Yes\" or \"No\".
 3. For dates, use the exact format requested.
 4. For numbers, use the exact number, no other format.
 5. For names, use the exact name as found in sources.
-6. If the question has an associated file, process it accordingly.
+6. If the question has an associated file, download the file first using the task ID.
 Examples of good responses:
 - \"42\"
+- \"Arturo Nunez\"
 - \"Yes\"
 - \"October 5, 2001\"
 - \"Buenos Aires\"
 Never include phrases like \"the answer is...\" or \"Based on my research\".
-Only return the exact answer."""
-
-# Define agent tools
-audio_tool = AudioTranscriptionTool()
-image_tool = ImageAnalysisTool()
-wikipedia_tool = WikipediaSearcher()
-
-tools = [audio_tool, image_tool, wikipedia_tool]
-
-# Define the custom agent using Dolphin model (free Mixtral)
-class MyAgent(CodeAgent):
-    def __init__(self):
-        model = HfApiModel(
-            model="cognitivecomputations/dolphin-2.6-mixtral-8x7b",
-            api_key=os.getenv("HF_API_TOKEN", "").strip(),
-            # No system_prompt here
-        )
-        super().__init__(model=model, tools=tools)
+Only return the exact answer.
+QUESTION:
+{question}
+"""
 
-    def __call__(self, question_dict):
-        system_message = {"role": "system", "content": SYSTEM_PROMPT}
-        user_message = {"role": "user", "content": question_dict.get("question", "")}
-        messages = [system_message, user_message]
-        # Pass messages directly in the call
-        return self.model(messages)
+        try:
+            result = self.agent.run(prompt)
+            print(f"Raw result from agent: {result}")
+
+            if isinstance(result, dict) and "answer" in result:
+                return str(result["answer"]).strip()
+            elif isinstance(result, str):
+                return result.strip()
+            elif isinstance(result, list):
+                for item in reversed(result):
+                    if isinstance(item, dict) and item.get("role") == "assistant" and "content" in item:
+                        return item["content"].strip()
+                return "ERROR: Unexpected list format"
+            else:
+                return "ERROR: Unexpected result type"
+        except Exception as e:
+            print(f"Exception during agent run: {e}")
+            return f"AGENT ERROR: {e}"
 
 # Evaluation + Submission function
 def run_and_submit_all(profile: gr.OAuthProfile | None):
@@ -70,7 +91,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
     submit_url = f"{api_url}/submit"
 
     try:
-        agent = MyAgent()
+        agent = GaiaAgent()
     except Exception as e:
         print(f"Error initializing agent: {e}")
         return f"Error initializing agent: {e}", None
@@ -97,7 +118,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         if not task_id:
             continue
         try:
-            submitted_answer = agent(item)
+            submitted_answer = agent(item.get("question", ""))
             answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
             results_log.append({
                 "Task ID": task_id,