Update app.py

app.py

@@ -34,7 +34,10 @@ def wikipedia_search(query: str) -> str:
         if not results:
             return "No Wikipedia results found."
         title = results[0]["title"]
-        summary_params = {"action": "query", "titles": title, "prop": "extracts", "exintro": True, "explaintext": True, "format": "json"}
+        summary_params = {
+            "action": "query", "titles": title, "prop": "extracts",
+            "exintro": True, "explaintext": True, "format": "json"
+        }
         summary_response = requests.get(search_url, params=summary_params, timeout=10)
         summary_data = summary_response.json()
         pages = summary_data.get("query", {}).get("pages", {})
@@ -66,53 +69,70 @@ class SmartAgent:
     def __init__(self):
         self.hf_token = os.getenv("HF_TOKEN")
         if not self.hf_token:
-            raise ValueError("HF_TOKEN environment variable not set!")
-        # Use HF Inference Providers (new format)
-        self.api_url = "https://api-inference.huggingface.co/v1/chat/completions"
-        self.model = "Qwen/Qwen2.5-72B-Instruct"
+            raise ValueError("HF_TOKEN not set!")
+        # mistralai/Mistral-7B-Instruct-v0.3 is free via HF Inference API (no credits needed)
+        self.api_url = "https://api-inference.huggingface.co/models/mistralai/Mistral-7B-Instruct-v0.3"
         self.headers = {
             "Authorization": f"Bearer {self.hf_token}",
             "Content-Type": "application/json"
         }
-        print(f"SmartAgent initialized with {self.model} (HF Inference API - FREE)")
+        print("SmartAgent initialized with Mistral-7B (FREE, no credits needed)")
 
-    def call_llm(self, messages: list) -> str:
+    def call_llm(self, prompt: str) -> str:
         payload = {
-            "model": self.model,
-            "messages": messages,
-            "max_tokens": 1024,
-            "temperature": 0.1,
+            "inputs": prompt,
+            "parameters": {
+                "max_new_tokens": 512,
+                "temperature": 0.1,
+                "return_full_text": False,
+                "stop": ["</s>", "[INST]"]
+            }
         }
         response = requests.post(self.api_url, headers=self.headers, json=payload, timeout=60)
         response.raise_for_status()
-        return response.json()["choices"][0]["message"]["content"]
+        result = response.json()
+        if isinstance(result, list) and len(result) > 0:
+            return result[0].get("generated_text", "").strip()
+        return str(result)
+
+    def build_prompt(self, messages: list) -> str:
+        """Build Mistral instruction format prompt"""
+        prompt = "<s>"
+        for msg in messages:
+            if msg["role"] == "user":
+                prompt += f"[INST] {msg['content']} [/INST]"
+            elif msg["role"] == "assistant":
+                prompt += f" {msg['content']}</s>"
+        return prompt
 
     def __call__(self, question: str) -> str:
         print(f"\nQuestion: {question[:100]}...")
 
-        system_prompt = """You are a precise AI assistant solving benchmark questions.
-
-You have access to these tools - call them by outputting exactly:
+        system = """You are a precise AI assistant solving benchmark questions.
+You can use these tools by outputting exactly:
 SEARCH: <query>
 WIKIPEDIA: <query>
 PYTHON: <code>
 
-Rules:
-- Use tools to look up facts you are unsure about
-- After gathering info, output your final answer as: ANSWER: <your answer>
-- Answer must be a number, short phrase, or comma-separated list
-- No articles (a/the), no units unless asked, no explanations
-- The answer is checked by exact match so be precise"""
+After gathering enough info, give your final answer as:
+ANSWER: <your answer>
+
+Rules for the answer:
+- Numbers only (no units unless asked)
+- Short phrases (no articles like a/the)
+- Comma-separated list if multiple items
+- Exact match required, be precise"""
 
         messages = [
-            {"role": "system", "content": system_prompt},
-            {"role": "user", "content": question}
+            {"role": "user", "content": f"{system}\n\nQuestion: {question}"}
         ]
 
-        for iteration in range(6):
-            response = self.call_llm(messages)
-            print(f"  LLM: {response[:200]}")
+        for iteration in range(5):
+            prompt = self.build_prompt(messages)
+            response = self.call_llm(prompt)
+            print(f"  LLM [{iteration}]: {response[:200]}")
 
+            # Check for final answer
             answer_match = re.search(r'ANSWER:\s*(.+?)(?:\n|$)', response, re.IGNORECASE)
             if answer_match:
                 answer = answer_match.group(1).strip()
@@ -125,13 +145,13 @@ Rules:
             if search_match:
                 query = search_match.group(1).strip()
                 print(f"  Tool: web_search({query})")
-                tool_result = f"Search results for '{query}':\n{web_search(query)}"
+                tool_result = f"Search results:\n{web_search(query)}"
 
             wiki_match = re.search(r'WIKIPEDIA:\s*(.+?)(?:\n|$)', response)
             if wiki_match:
                 query = wiki_match.group(1).strip()
                 print(f"  Tool: wikipedia({query})")
-                tool_result = f"Wikipedia results for '{query}':\n{wikipedia_search(query)}"
+                tool_result = f"Wikipedia:\n{wikipedia_search(query)}"
 
             python_match = re.search(r'PYTHON:\s*```(?:python)?\n?(.*?)```', response, re.DOTALL)
             if not python_match:
@@ -143,16 +163,18 @@ Rules:
 
             if tool_result:
                 messages.append({"role": "assistant", "content": response})
-                messages.append({"role": "user", "content": tool_result})
+                messages.append({"role": "user", "content": f"{tool_result}\n\nNow provide your ANSWER: <answer>"})
             else:
                 messages.append({"role": "assistant", "content": response})
-                messages.append({"role": "user", "content": "Please provide your final answer now using: ANSWER: <your answer>"})
+                messages.append({"role": "user", "content": "Provide your final answer as: ANSWER: <answer>"})
 
-        last_response = self.call_llm(messages)
-        answer_match = re.search(r'ANSWER:\s*(.+?)(?:\n|$)', last_response, re.IGNORECASE)
+        # Final attempt
+        messages.append({"role": "user", "content": "Give only the final answer as: ANSWER: <answer>"})
+        last = self.call_llm(self.build_prompt(messages))
+        answer_match = re.search(r'ANSWER:\s*(.+?)(?:\n|$)', last, re.IGNORECASE)
         if answer_match:
             return answer_match.group(1).strip()
-        return last_response.strip().split('\n')[0][:200]
+        return last.strip().split('\n')[0][:200]
 
 
 def run_and_submit_all(profile: gr.OAuthProfile | None):
@@ -232,7 +254,7 @@ with gr.Blocks() as demo:
     gr.Markdown("# 🤖 Smart Agent — GAIA Benchmark Runner")
     gr.Markdown(
         """
-        **Powered by Qwen2.5-72B via HuggingFace Inference API (FREE)**
+        **Powered by Mistral-7B via HuggingFace Inference API (100% FREE - no credits needed)**
 
         **Instructions:**
         1. Make sure `HF_TOKEN` is set in your Space secrets