Update app.py

app.py

@@ -9,125 +9,110 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
 class BasicAgent:
     def __init__(self):
-        # 1. Initialize the Model (the 'brain')
-        # This wrapper is the most stable version for HF Inference API in 2026.
-        # It will automatically use your HF_TOKEN secret if added to the Space.
+        # We use Qwen2.5-72B for its strong reasoning and code generation
         self.model = InferenceClientModel(model_id="Qwen/Qwen2.5-72B-Instruct")
         
-        # 2. Initialize the CodeAgent (the 'body')
-        # - tools=[]: We start with no external tools.
-        # - add_base_tools=False: This prevents the 'ddgs' / DuckDuckGo error.
-        # Note: CodeAgent still has a built-in Python interpreter to solve math/logic!
+        # CodeAgent natively handles Python code execution
         self.agent = CodeAgent(
             tools=[], 
             model=self.model,
             add_base_tools=False 
         )
-        print("Agent successfully initialized with Python Interpreter (No ddgs needed).")
 
     def __call__(self, question: str) -> str:
-        # 3. Prompting for Exact Match scoring
-        # We tell the agent to be as direct as possible.
-        clean_prompt = (
-            f"Question: {question}\n\n"
-            "Final Answer Requirement: Provide ONLY the numeric or text value of the answer. "
-            "Do not include any explanation, units, or 'The answer is' text. "
-            "Do not include 'FINAL ANSWER' in your output."
+        # This prompt is the 'gold standard' for GAIA exact-match scoring
+        strict_prompt = (
+            f"You are a general AI assistant. Solve this task: {question}\n\n"
+            "Report your thoughts, and finish your answer with the following template: "
+            "FINAL ANSWER: [YOUR FINAL ANSWER]\n\n"
+            "YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list. "
+            "If the answer is a number, do not use units ($, %, kg). "
+            "If the answer is a string, don't use articles (a, an, the) or abbreviations."
         )
         
         try:
-            # The agent will write and run Python code if the question requires it.
-            result = self.agent.run(clean_prompt)
-            return str(result).strip()
+            # Run the agent
+            raw_result = self.agent.run(strict_prompt)
+            
+            # --- Strict Post-Processing ---
+            # Extract only the content after 'FINAL ANSWER:' if it exists
+            result_str = str(raw_result)
+            if "FINAL ANSWER:" in result_str:
+                result_str = result_str.split("FINAL ANSWER:")[-1]
+            
+            # Remove markdown, quotes, and trailing punctuation
+            clean_ans = result_str.replace("`", "").replace('"', "").replace("'", "").strip()
+            return clean_ans.rstrip('.')
         except Exception as e:
-            print(f"Error during agent execution: {e}")
-            return "Error solving question"
+            print(f"Agent Error: {e}")
+            return "Error"
 
 def run_and_submit_all(profile: gr.OAuthProfile | None):
-    space_id = os.getenv("SPACE_ID")
-
     if not profile:
         return "Please Login to Hugging Face with the button.", None
 
-    username = f"{profile.username}"
-    api_url = DEFAULT_API_URL
-    questions_url = f"{api_url}/questions"
-    submit_url = f"{api_url}/submit"
-    
-    # URL to your code for verification
-    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main" if space_id else "https://huggingface.co/spaces"
+    username = profile.username
+    space_id = os.getenv("SPACE_ID")
+    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
 
     # 1. Fetch Questions
     try:
-        response = requests.get(questions_url, timeout=15)
-        response.raise_for_status()
+        response = requests.get(f"{DEFAULT_API_URL}/questions", timeout=15)
         questions_data = response.json()
     except Exception as e:
-        return f"Error fetching questions: {e}", None
+        return f"Fetch Error: {e}", None
 
-    # 2. Run Agent
-    # Instantiate inside the function to ensure a fresh session
-    try:
-        agent = BasicAgent()
-    except Exception as e:
-        return f"Error initializing agent: {e}", None
-
-    results_log = []
+    # 2. Setup Agent
+    agent = BasicAgent()
     answers_payload = []
-    
+    results_log = []
+
     print(f"Starting evaluation for {len(questions_data)} questions...")
 
     for item in questions_data:
         task_id = item.get("task_id")
         question_text = item.get("question")
         
-        if not task_id or question_text is None:
-            continue
-
+        # --- File Download Logic ---
+        # Some GAIA questions refer to a file. We download it so the Python tool can read it.
+        file_url = f"{DEFAULT_API_URL}/files/{task_id}"
         try:
-            submitted_answer = agent(question_text)
-            answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
-            results_log.append({"Task ID": task_id, "Question": question_text[:80], "Submitted Answer": submitted_answer})
-        except Exception as e:
-            results_log.append({"Task ID": task_id, "Question": question_text[:80], "Submitted Answer": f"ERROR: {e}"})
-
-    # 3. Submit Results
-    submission_data = {
-        "username": username.strip(), 
-        "agent_code": agent_code, 
-        "answers": answers_payload
-    }
+            file_res = requests.get(file_url, timeout=5)
+            if file_res.status_code == 200:
+                # Save the file with its task_id as the name
+                with open(task_id, "wb") as f:
+                    f.write(file_res.content)
+                question_text += f"\n\n[System Note: A file for this task has been downloaded to your directory as '{task_id}'. Use Python to read/analyze it.]"
+        except:
+            pass # No file for this task
+
+        # 3. Generate Answer
+        ans = agent(question_text)
+        
+        answers_payload.append({"task_id": task_id, "submitted_answer": ans})
+        results_log.append({"Task ID": task_id, "Answer": ans})
 
+    # 4. Submit
     try:
-        response = requests.post(submit_url, json=submission_data, timeout=60)
-        response.raise_for_status()
-        result_data = response.json()
-        
-        final_status = (
-            f"Submission Successful!\n"
-            f"User: {result_data.get('username')}\n"
-            f"Overall Score: {result_data.get('score', 0)}% "
-            f"({result_data.get('correct_count', 0)}/{result_data.get('total_attempted', 0)} correct)"
-        )
-        return final_status, pd.DataFrame(results_log)
+        sub_res = requests.post(f"{DEFAULT_API_URL}/submit", json={
+            "username": username,
+            "agent_code": agent_code,
+            "answers": answers_payload
+        })
+        data = sub_res.json()
+        status = f"Success! Score: {data.get('score')}% ({data.get('correct_count')}/{data.get('total_attempted')})"
+        return status, pd.DataFrame(results_log)
     except Exception as e:
         return f"Submission Failed: {e}", pd.DataFrame(results_log)
 
 # --- Gradio UI ---
 with gr.Blocks() as demo:
-    gr.Markdown("# GAIA Final Evaluation Solver")
-    gr.Markdown("Click 'Login' then 'Run' to solve all questions and submit your score.")
-
+    gr.Markdown("# GAIA Pro Solver")
     gr.LoginButton()
-    run_button = gr.Button("Run Evaluation & Submit All Answers", variant="primary")
-
-    status_output = gr.Textbox(label="Submission Status", lines=5)
-    results_table = gr.DataFrame(label="Agent Answers Log", wrap=True)
-
-    run_button.click(
-        fn=run_and_submit_all,
-        outputs=[status_output, results_table]
-    )
+    run_btn = gr.Button("Run & Submit", variant="primary")
+    status = gr.Textbox(label="Result")
+    table = gr.DataFrame(label="Log")
+    run_btn.click(fn=run_and_submit_all, outputs=[status, table])
 
 if __name__ == "__main__":
     demo.launch()