Update app.py

app.py

@@ -5,7 +5,6 @@ import requests
 import pandas as pd
 from smolagents import (
     CodeAgent,
-    LiteLLMModel,
     InferenceClientModel,
     DuckDuckGoSearchTool,
     WikipediaSearchTool,
@@ -16,117 +15,64 @@ from smolagents import (
 
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
-# --- Custom Throttled Model to protect Gemini ---
-class ThrottledGeminiModel(LiteLLMModel):
-    def __call__(self, *args, **kwargs):
-        time.sleep(5) # Base 5-second delay to stay under 15 RPM
-        return super().__call__(*args, **kwargs)
-
 @tool
 def get_current_date_time() -> str:
     """Returns the current date and time in ISO format."""
     from datetime import datetime
     return datetime.now().isoformat()
 
-class FailproofAgent:
+class StrictHuggingFaceAgent:
     def __init__(self):
-        print("Initializing Failproof Cascade Agent...")
-        self.models = []
+        print("Initializing Strict Hugging Face Agent with Few-Shot Prompting...")
         
-        # 1. Primary: Gemini 2.0 Flash (1500 daily requests, huge context)
-        gemini_key = os.getenv("GEMINI_API_KEY")
-        if gemini_key:
-            self.models.append({
-                "name": "Gemini 2.0 Flash",
-                "model": ThrottledGeminiModel(model_id="gemini/gemini-2.0-flash", api_key=gemini_key)
-            })
-            
-        # 2. Secondary: HF Qwen2.5-Coder (Great for code, serverless)
-        hf_token = os.getenv("HF_TOKEN") or os.getenv("HF_TOKEN")
-        if hf_token:
-            self.models.append({
-                "name": "Hugging Face Qwen2.5 Coder",
-                "model": InferenceClientModel(model_id="Qwen/Qwen2.5-Coder-32B-Instruct", token=hf_token)
-            })
-            
-        # 3. Tertiary: Groq Llama 3.3 (Fast, smart fallback)
-        groq_key = os.getenv("GROQ_API_KEY")
-        if groq_key:
-            self.models.append({
-                "name": "Groq Llama 3.3 70B",
-                "model": LiteLLMModel(model_id="groq/llama-3.3-70b-versatile", api_key=groq_key)
-            })
+        hf_token = os.getenv("HF_TOKEN")
+        if not hf_token:
+            raise ValueError("HF_TOKEN environment variable not set in Space Secrets.")
             
-        # 4. Emergency: OpenRouter Free Pool (Decentralized backup)
-        or_key = os.getenv("OPENROUTER_API_KEY")
-        if or_key:
-            self.models.append({
-                "name": "OpenRouter Auto-Free",
-                "model": LiteLLMModel(model_id="openrouter/openrouter/free", api_key=or_key)
-            })
-            
-        if not self.models:
-            raise ValueError("No API keys found! Please set at least one in Space Secrets.")
-            
-        self.current_model_idx = 0
+        self.model = InferenceClientModel(
+            model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
+            token=hf_token,
+        )
+        
         self.tools = [
             DuckDuckGoSearchTool(),
             WikipediaSearchTool(),
+            VisitWebpageTool(), 
             PythonInterpreterTool(),
-            VisitWebpageTool(), # Allows the agent to read inside websites
             get_current_date_time,
         ]
-        print(f"Agent armed with {len(self.models)} fallback brains. Ready to go.")
+        
+        self.agent = CodeAgent(
+            tools=self.tools,
+            model=self.model,
+            max_steps=7,
+            additional_authorized_imports=["datetime", "re", "json", "math", "collections", "pandas", "requests", "bs4"],
+        )
+        print("Agent ready.")
 
     def __call__(self, question: str) -> str:
         print(f"\nAgent received question: {question[:80]}...")
-        max_retries_per_model = 3
+        max_retries = 3
         
-        # Keep trying models until we run out of backups
-        while self.current_model_idx < len(self.models):
-            current_brain = self.models[self.current_model_idx]
-            print(f"🧠 USING BRAIN: {current_brain['name']}")
-            
-            # Re-instantiate the agent cleanly for this attempt
-            agent = CodeAgent(
-                tools=self.tools,
-                model=current_brain["model"],
-                max_steps=7,
-                additional_authorized_imports=["datetime", "re", "json", "math", "collections", "pandas", "requests"],
-            )
-            
-            for attempt in range(max_retries_per_model):
-                try:
-                    time.sleep(2) 
-                    answer = agent.run(question)
-                    print(f"Agent answer: {str(answer)[:200]}")
-                    return str(answer)
-                except Exception as e:
-                    err_msg = str(e).lower()
-                    print(f"⚠️ Agent Error: {err_msg}")
+        for attempt in range(max_retries):
+            try:
+                time.sleep(2) 
+                answer = self.agent.run(question)
+                # Clean up any accidental leading/trailing whitespace or quotes the agent might slip in
+                clean_answer = str(answer).strip(" '\"\n\t.") 
+                print(f"Agent answer: {clean_answer}")
+                return clean_answer
+            except Exception as e:
+                err_msg = str(e).lower()
+                if "429" in err_msg or "rate limit" in err_msg or "too many requests" in err_msg:
+                    wait_time = 20 * (attempt + 1)
+                    print(f"Rate limit hit! Pausing for {wait_time} seconds before retrying...")
+                    time.sleep(wait_time)
+                else:
+                    print(f"Agent error processing question: {e}")
+                    return f"Error: {str(e)}"
                     
-                    # FATAL QUOTA ERROR: Break the retry loop and switch brains
-                    if "402" in err_msg or "payment required" in err_msg or "quota" in err_msg or "limit 0" in err_msg or "spend limit" in err_msg:
-                        print(f"🚨 FATAL QUOTA HIT on {current_brain['name']}. Swapping to backup brain...")
-                        break # This exits the attempt loop and moves to the next model
-                        
-                    # TEMPORARY RATE LIMIT: Pause and retry the same brain
-                    elif "429" in err_msg or "rate limit" in err_msg or "too many requests" in err_msg:
-                        wait_time = 20 * (attempt + 1)
-                        print(f"⏳ Temporary rate limit. Pausing for {wait_time}s...")
-                        time.sleep(wait_time)
-                        continue
-                        
-                    # OTHER ERRORS (Code failures, etc): Retry
-                    else:
-                        print("Retrying due to generic error...")
-                        continue
-                        
-            # If we exit the loop, this brain has failed completely. Move to the next one.
-            print(f"⏭️ Exhausted retries or hit hard limit on {current_brain['name']}. Escalating...")
-            self.current_model_idx += 1
-            
-        return "Error: All available models exhausted their quotas or failed."
+        return "Error: Rate limit exceeded after maximum retries."
 
 # --- App Runner ---
 def run_and_submit_all(profile: gr.OAuthProfile | None):
@@ -143,7 +89,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
     submit_url = f"{api_url}/submit"
     
     try:
-        agent = FailproofAgent()
+        agent = StrictHuggingFaceAgent()
     except Exception as e:
         print(f"Error instantiating agent: {e}")
         return f"Error initializing agent: {e}", None
@@ -171,29 +117,40 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
         if not task_id or not question_text:
             continue
             
+        # Inject the file URL if it exists
         if file_url:
             question_text += f"\n\n[IMPORTANT: This task requires analyzing an attached file. You MUST download or read it directly from this URL: {file_url} using your Python tool.]"
             
-        strict_prompt = (
+        # The ultimate, unbreakable strict prompt WITH few-shot examples
+        ultra_strict_prompt = (
             f"{question_text}\n\n"
-            "CRITICAL SUBMISSION INSTRUCTIONS:\n"
-            "The system evaluating your answer is a strict automated parser.\n"
-            "1. You MUST output ONLY the final requested answer.\n"
-            "2. DO NOT include any conversational text, explanations, or reasoning in your final output.\n"
-            "3. If the answer is a name, number, or short string, output ONLY that exact string.\n"
-            "4. For numbers, do not include symbols unless explicitly requested."
-            "5. **ULTRATHINK** and double check the response making sure the return answer."
+            "=== CRITICAL OUTPUT INSTRUCTIONS ===\n"
+            "You are being evaluated by a strict programmatic regex parser.\n"
+            "Your final answer MUST consist of ONLY the exact requested name, number, or string.\n"
+            "DO NOT wrap your answer in quotes, DO NOT add a trailing period, and DO NOT provide any explanation or conversational filler.\n\n"
+            "Here are examples of perfect submissions:\n"
+            "Example 1\n"
+            "Question: What is the first name of the only Malko Competition recipient from the 20th Century (after 1977) whose nationality on record is a country that no longer exists?\n"
+            "Answer: Vladimir\n\n"
+            "Example 2\n"
+            "Question: How many at bats did the Yankee with the most walks in the 1977 regular season have that same season?\n"
+            "Answer: 519\n\n"
+            "Example 3\n"
+            "Question: .rewsna eht sa \"tfel\" drow eht fo etisoppo eht etirw ,ecnetnes siht dnatsrednu uoy fI\n"
+            "Answer: right\n\n"
+            "Failure to follow these instructions perfectly will result in an immediate score of 0."
         )
             
         try:
-            submitted_answer = agent(strict_prompt)
+            submitted_answer = agent(ultra_strict_prompt)
             answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
         except Exception as e:
             results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"ERROR: {e}"})
             
-        print("Cooling down for 10 seconds to protect quotas...")
-        time.sleep(10)
+        # 15 second cooldown to protect your new Hugging Face token limits
+        print("Cooling down for 15 seconds to protect quotas...")
+        time.sleep(15)
             
     if not answers_payload:
         return "No answers.", pd.DataFrame(results_log)
@@ -219,14 +176,13 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
 
 # --- Build Gradio UI ---
 with gr.Blocks() as demo:
-    gr.Markdown("# The Failproof Multi-Model Agent Runner")
+    gr.Markdown("# Strict Hugging Face Evaluation Runner (Few-Shot Edition)")
     gr.Markdown(
         """
         **Instructions:**
-        1. Ensure your API keys (`GEMINI_API_KEY`, `NEW_HF_TOKEN`, `GROQ_API_KEY`, etc.) are set in Space Secrets.
+        1. Ensure your fresh `HF_TOKEN` is set in Space Secrets.
         2. Log in below.
         3. Click 'Run Evaluation & Submit' to start. 
-        *(Watch the logs! If a model dies, it will automatically hot-swap to the next one).*
         """
     )
     gr.LoginButton()