Update app.py

app.py

@@ -1,11 +1,12 @@
 import os
-import re
 import gradio as gr
 import requests
 import pandas as pd
+import re
 
 from huggingface_hub import InferenceClient
-from smolagents import CodeAgent, DuckDuckGoSearchTool, HfApiModel
+
+from smolagents import CodeAgent, DuckDuckGoSearchTool, InferenceClientModel
 
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
@@ -71,56 +72,63 @@ SYSTEM_PROMPT = (
 
 class BasicAgent:
     """
-    Agente usando smolagents + HfApiModel + DuckDuckGoSearchTool,
-    ajustado para o cenário do GAIA (EXACT MATCH).
+    Agente simples que usa diretamente o InferenceClient do Hugging Face
+    para responder as questões do GAIA.
+
+    Não usa ferramentas externas (search), mas é MUITO mais estável no Space.
     """
 
     def __init__(self):
-        print("Initializing GAIA agent with HfApiModel + DuckDuckGoSearchTool...")
+        print("Initializing Simple GAIA Agent with InferenceClient...")
 
-        # ferramenta de busca da própria smolagents (usa ddgs por baixo)
-        self.search_tool = DuckDuckGoSearchTool()
+        hf_token = os.getenv("HF_TOKEN")
+        if not hf_token:
+            raise ValueError(
+                "HF_TOKEN not found! Configure it como Secret em Settings → Variables."
+            )
 
-        # Modelo chamado via API do Hugging Face
-        # Você pode trocar o model_id se o curso recomendar outro
-        self.model = HfApiModel(
-            model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
-            max_new_tokens=256,
-            temperature=0.2,
+        # escolha de modelo (pode trocar por outro compatível com Inference API)
+        self.client = InferenceClient(
+            model="Qwen/Qwen2.5-72B-Instruct",
+            token=hf_token,
         )
 
-        # CodeAgent: agente que consegue chamar tools e escrever código
-        self.agent = CodeAgent(
-            model=self.model,
-            tools=[self.search_tool],
-            add_base_tools=False,  # deixamos só a search tool pra ficar simples
+        # prompt base com regras de EXACT MATCH
+        self.system_instructions = (
+            "You are solving GAIA benchmark questions.\n"
+            "Rules:\n"
+            "- Answer ONLY with the final answer.\n"
+            "- No explanations, no reasoning, no extra words.\n"
+            "- Do NOT write 'Final answer', 'Answer:', etc.\n"
+            "- If the answer is a number, output just the number.\n"
+            "- Your output will be compared with EXACT MATCH.\n"
         )
 
     def __call__(self, question: str) -> str:
-        print(f"Processing question: {question[:80]}...")
+        print(f"\n=== NEW QUESTION ===\n{question}\n")
 
-        # Instruções de EXACT MATCH embutidas no prompt
         prompt = (
-            "You are solving GAIA Level 1 questions.\n"
-            "You have access to a web search tool and MUST use it whenever needed "
-            "to obtain accurate, up-to-date information.\n"
-            "RULES:\n"
-            " - Return ONLY the final answer as a short string.\n"
-            " - NO explanations.\n"
-            " - NO 'Final answer', 'Answer:', etc.\n"
-            " - If the answer is a number, output just the number.\n"
-            "Question:\n"
-            f"{question}\n"
+            self.system_instructions
+            + "\nQuestion:\n"
+            + question
+            + "\n\nAnswer (remember: ONLY the final answer):"
         )
 
         try:
-            raw = self.agent.run(prompt)
+            # usamos text_generation para evitar problemas com chat_completion
+            raw = self.client.text_generation(
+                prompt,
+                max_new_tokens=64,
+                temperature=0.1,
+                top_p=0.9,
+                stop_sequences=["\n"],  # para não vir um parágrafo gigante
+            )
+            print("RAW MODEL OUTPUT:", repr(raw))
             final = clean_answer(raw)
-            print(f"Raw answer: {raw!r}")
-            print(f"Final cleaned answer: {final!r}")
+            print("CLEANED ANSWER:", repr(final))
             return final
         except Exception as e:
-            print(f"Error while running agent: {e}")
+            print("ERROR calling InferenceClient:", e)
             return ""
 
 # =========================================================