Update agent.py

agent.py

@@ -1,4 +1,5 @@
 import os
+import re
 from smolagents import CodeAgent, DuckDuckGoSearchTool, tool, InferenceClientModel
 
 @tool
@@ -9,41 +10,58 @@ def calculate_math(expression: str) -> str:
     """
     try:
         allowed_chars = "0123456789+-*/(). "
-        if all(c in allowed_chars for c in expression):
-            return str(eval(expression))
+        cleaned = "".join(c for c in expression if c in allowed_chars)
+        if cleaned:
+            return str(eval(cleaned))
         return "Error: Invalid characters in math expression."
     except Exception as e:
         return f"Error executing calculation: {str(e)}"
 
 class CourseEvaluationAgent:
     def __init__(self):
-        print("[Agent System] Booting Engine...")
+        print("[Agent System] Booting Upgraded Engine...")
         
-        # In this environment version, InferenceClientModel handles the Hub implementation.
-        # It reads your HF_TOKEN secret natively behind the scenes.
+        # Use the highly capable Llama 3.3 70B model via serverless client
         self.model = InferenceClientModel(model_id="meta-llama/Llama-3.3-70B-Instruct")
         
-        # Pass the fully instantiated concrete model to the model positional argument
+        # Initialize the CodeAgent
         self.agent = CodeAgent(
             tools=[DuckDuckGoSearchTool(), calculate_math], 
             model=self.model,
-            max_steps=5,      # Prevents execution hang ups or infinite loops
+            max_steps=5,
             verbosity_level=1
         )
         
-        extra_instructions = (
-            "\nCRITICAL INSTRUCTION: Be concise. When responding with your final answer, "
-            "provide just the factual outcome or specific answer requested. "
-            "Do not include unnecessary conversational filler."
+        # STRICT PROMPT FIX: Completely override the prompt system template 
+        # to guarantee the agent ONLY outputs the clean, final answer.
+        self.agent.prompt_templates["system_prompt"] = (
+            "You are an expert factual agent. You have access to a python code interpreter and web search.\n"
+            "CRITICAL RULES:\n"
+            "1. Run code or search to find the exact truth.\n"
+            "2. Your final response MUST contain ONLY the direct, raw answer text (e.g., '42', 'Paris', 'John Doe').\n"
+            "3. Absolutely DO NOT include conversational text like 'The answer is...', 'According to search...', or punctuation filler.\n"
+            "4. Be completely literal."
         )
-        
-        if "system_prompt" in self.agent.prompt_templates:
-            self.agent.prompt_templates["system_prompt"] += extra_instructions
 
     def run_task(self, question: str) -> str:
         try:
+            # Let the agent think and execute tools
             output = self.agent.run(question)
-            return str(output).strip()
+            final_ans = str(output).strip()
+            
+            # POST-PROCESSING CLEANER: If the model still outputs conversational fluff, 
+            # we programmaticly clean common phrases before returning to the grader.
+            clean_patterns = [
+                r"^[Tt]he answer is\s*[:\-]?\s*",
+                r"^[Ff]inal [Aa]nswer\s*[:\-]?\s*",
+                r"^[\"']",  # Strip starting quotes
+                r"[\"']$"   # Strip ending quotes
+            ]
+            for pattern in clean_patterns:
+                final_ans = re.sub(pattern, "", final_ans)
+                
+            return final_ans.strip()
+            
         except Exception as e:
             print(f"[Agent Error] Failed execution: {e}")
-            return f"Error: {str(e)[:50]}"
+            return "Error calculating answer."