Update app.py

app.py

@@ -1,13 +1,11 @@
 import os
-import gradio as gr
+import re
 import requests
 import pandas as pd
-import re
+import gradio as gr
 
 from huggingface_hub import InferenceClient
 
-from smolagents import CodeAgent, DuckDuckGoSearchTool, InferenceClientModel
-
 # --- Constants ---
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
@@ -72,24 +70,22 @@ SYSTEM_PROMPT = (
 
 class BasicAgent:
     """
-    Agente simples que usa diretamente o InferenceClient do Hugging Face
-    para responder as questões do GAIA.
-
-    Não usa ferramentas externas (search), mas é estável no Space.
+    Agente simples usando InferenceClient.chat_completion
+    para responder as questões do GAIA em modo conversacional.
     """
 
     def __init__(self):
-        print("Initializing Simple GAIA Agent with InferenceClient...")
+        print("Initializing Simple GAIA Agent with chat_completion...")
 
         hf_token = os.getenv("HF_TOKEN")
         if not hf_token:
             raise ValueError(
-                "HF_TOKEN not found! Configure it como Secret em Settings → Variables."
+                "HF_TOKEN not found! Crie um Secret chamado HF_TOKEN em Settings → Variables."
             )
 
-        # 🔹 Modelo que suporta text-generation via Inference API
+        # Modelo que sabemos ser suportado como 'conversational'
         self.client = InferenceClient(
-            model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+            model="Qwen/Qwen2.5-72B-Instruct",  # o mesmo que a infra mostrou no log
             token=hf_token,
         )
 
@@ -106,28 +102,40 @@ class BasicAgent:
     def __call__(self, question: str) -> str:
         print(f"\n=== NEW QUESTION ===\n{question}\n")
 
-        prompt = (
-            self.system_instructions
-            + "\nQuestion:\n"
-            + question
-            + "\n\nAnswer (remember: ONLY the final answer):"
-        )
+        messages = [
+            {"role": "system", "content": self.system_instructions},
+            {
+                "role": "user",
+                "content": (
+                    question
+                    + "\n\nRemember: reply ONLY with the final answer, nothing else."
+                ),
+            },
+        ]
 
         try:
-            # usamos text_generation para evitar problemas com chat_completion
-            raw = self.client.text_generation(
-                prompt,
-                max_new_tokens=64,
+            completion = self.client.chat_completion(
+                messages=messages,
+                max_tokens=64,
                 temperature=0.1,
                 top_p=0.9,
-                stop_sequences=["\n"],  # para não vir um parágrafo gigante
             )
+
+            # compatível com os dois formatos (.message["content"] ou .message.content)
+            choice = completion.choices[0]
+            message = choice.message
+            if isinstance(message, dict):
+                raw = message.get("content", "")
+            else:
+                raw = getattr(message, "content", "")
+
             print("RAW MODEL OUTPUT:", repr(raw))
             final = clean_answer(raw)
             print("CLEANED ANSWER:", repr(final))
             return final
+
         except Exception as e:
-            print("ERROR calling InferenceClient:", e)
+            print("ERROR calling InferenceClient.chat_completion:", e)
             return ""
 
 # =========================================================