Update app.py

app.py

@@ -67,48 +67,31 @@ def run_python(code: str) -> str:
 
 class SmartAgent:
     def __init__(self):
-        self.hf_token = os.getenv("HF_TOKEN")
-        if not self.hf_token:
-            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("SmartAgent initialized with Mistral-7B (FREE, no credits needed)")
+        self.api_key = os.getenv("GEMINI_API_KEY")
+        if not self.api_key:
+            raise ValueError("GEMINI_API_KEY environment variable not set!")
+        # Gemini 1.5 Flash - FREE tier: 1500 requests/day, 1M tokens/min
+        self.api_url = f"https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key={self.api_key}"
+        print("SmartAgent initialized with Gemini 1.5 Flash (FREE - 1500 req/day)")
 
     def call_llm(self, prompt: str) -> str:
         payload = {
-            "inputs": prompt,
-            "parameters": {
-                "max_new_tokens": 512,
+            "contents": [{"parts": [{"text": prompt}]}],
+            "generationConfig": {
                 "temperature": 0.1,
-                "return_full_text": False,
-                "stop": ["</s>", "[INST]"]
+                "maxOutputTokens": 1024,
             }
         }
-        response = requests.post(self.api_url, headers=self.headers, json=payload, timeout=60)
+        response = requests.post(self.api_url, json=payload, timeout=60)
         response.raise_for_status()
-        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
+        data = response.json()
+        return data["candidates"][0]["content"]["parts"][0]["text"].strip()
 
     def __call__(self, question: str) -> str:
         print(f"\nQuestion: {question[:100]}...")
 
         system = """You are a precise AI assistant solving benchmark questions.
+
 You can use these tools by outputting exactly:
 SEARCH: <query>
 WIKIPEDIA: <query>
@@ -118,18 +101,15 @@ 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": "user", "content": f"{system}\n\nQuestion: {question}"}
-        ]
-
-        for iteration in range(5):
-            prompt = self.build_prompt(messages)
-            response = self.call_llm(prompt)
+- Numbers only (no units unless asked, no commas in numbers)
+- Short phrases (no articles like a/the, no abbreviations for proper nouns)
+- Comma-separated list if multiple items needed
+- Exact match required - be very precise"""
+
+        conversation = f"{system}\n\nQuestion: {question}"
+
+        for iteration in range(6):
+            response = self.call_llm(conversation)
             print(f"  LLM [{iteration}]: {response[:200]}")
 
             # Check for final answer
@@ -145,13 +125,13 @@ Rules for the answer:
             if search_match:
                 query = search_match.group(1).strip()
                 print(f"  Tool: web_search({query})")
-                tool_result = f"Search results:\n{web_search(query)}"
+                tool_result = f"Search results for '{query}':\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:\n{wikipedia_search(query)}"
+                tool_result = f"Wikipedia results for '{query}':\n{wikipedia_search(query)}"
 
             python_match = re.search(r'PYTHON:\s*```(?:python)?\n?(.*?)```', response, re.DOTALL)
             if not python_match:
@@ -162,15 +142,13 @@ Rules for the answer:
                 tool_result = f"Python output:\n{run_python(code)}"
 
             if tool_result:
-                messages.append({"role": "assistant", "content": response})
-                messages.append({"role": "user", "content": f"{tool_result}\n\nNow provide your ANSWER: <answer>"})
+                conversation += f"\n\nAssistant: {response}\n\nTool Result: {tool_result}\n\nNow provide your ANSWER: <answer>"
             else:
-                messages.append({"role": "assistant", "content": response})
-                messages.append({"role": "user", "content": "Provide your final answer as: ANSWER: <answer>"})
+                conversation += f"\n\nAssistant: {response}\n\nProvide your final answer as: ANSWER: <answer>"
 
         # Final attempt
-        messages.append({"role": "user", "content": "Give only the final answer as: ANSWER: <answer>"})
-        last = self.call_llm(self.build_prompt(messages))
+        conversation += "\n\nGive only the final answer as: ANSWER: <answer>"
+        last = self.call_llm(conversation)
         answer_match = re.search(r'ANSWER:\s*(.+?)(?:\n|$)', last, re.IGNORECASE)
         if answer_match:
             return answer_match.group(1).strip()
@@ -254,10 +232,10 @@ with gr.Blocks() as demo:
     gr.Markdown("# 🤖 Smart Agent — GAIA Benchmark Runner")
     gr.Markdown(
         """
-        **Powered by Mistral-7B via HuggingFace Inference API (100% FREE - no credits needed)**
+        **Powered by Google Gemini 1.5 Flash (FREE - 1500 requests/day)**
 
         **Instructions:**
-        1. Make sure `HF_TOKEN` is set in your Space secrets
+        1. Make sure `GEMINI_API_KEY` is set in your Space secrets
         2. Log in with your Hugging Face account below
         3. Click **Run Evaluation & Submit All Answers**
         """