Update app.py

app.py

@@ -1,192 +1,278 @@
 import os
+
 import gradio as gr
-import requests
 import pandas as pd
+import requests
 from smolagents import CodeAgent, DuckDuckGoSearchTool, LiteLLMModel
 
-# --- Constants ---
+
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+RESULT_COLUMNS = ["Task ID", "Question", "Submitted Answer"]
+
 
-# --- Agent Definition ---
 class BasicAgent:
     def __init__(self):
-        print("Inicializando o Agente do GAIA...")
-        
-        # Define o modelo
-        self.model = LiteLLMModel(model_id="huggingface/Qwen/Qwen2.5-Coder-32B-Instruct")
-        
-        # Define as ferramentas
-        self.tools = [DuckDuckGoSearchTool()]
-        
-        # Prompt customizado para forçar o formato EXACT MATCH do GAIA
-        custom_prompt = """
-        You are an expert AI assistant solving tasks from the GAIA benchmark.
-        Your final answer MUST be extremely concise and exact. 
-        Do NOT include any conversational text, explanations, or the words "FINAL ANSWER" in your final output.
-        If the question asks for a comma-separated list, provide ONLY the list.
-        If the question asks for a number, provide ONLY the number.
-        """
-        
-        # Instancia o agente
+        print("Inicializando o agente GAIA...")
+
+        self.model = LiteLLMModel(
+            model_id="huggingface/Qwen/Qwen2.5-Coder-32B-Instruct"
+        )
         self.agent = CodeAgent(
-            tools=self.tools,
+            tools=[DuckDuckGoSearchTool()],
             model=self.model,
             max_steps=6,
-            description="Agent designed to solve GAIA benchmark questions with exact match answers."
+            description=(
+                "Agent designed to solve GAIA benchmark questions with "
+                "exact-match answers."
+            ),
+        )
+
+        exact_match_prompt = """
+You are an expert AI assistant solving tasks from the GAIA benchmark.
+Return only the requested final answer.
+Do not add explanations, conversational text, markdown, or the words
+"FINAL ANSWER".
+If a comma-separated list is requested, return only that list.
+If a number is requested, return only that number.
+"""
+        self.agent.system_prompt = (
+            exact_match_prompt.strip() + "\n\n" + self.agent.system_prompt
         )
-        
-        # Injetando a instrução no sistema
-        self.agent.system_prompt = custom_prompt + "\n" + self.agent.system_prompt
 
     def __call__(self, question: str) -> str:
-        print(f"Agent received question (first 50 chars): {question[:50]}...")
-        try:
-            # O agente executa a pesquisa e raciocina
-            resposta_final = self.agent.run(question)
-            
-            # Limpeza básica de segurança para evitar que a string "FINAL ANSWER" vaze
-            resposta_limpa = str(resposta_final).replace("FINAL ANSWER", "").strip()
-            
-            print(f"Agent returning answer: {resposta_limpa}")
-            return resposta_limpa
-        except Exception as e:
-            print(f"Erro durante o raciocínio do agente: {e}")
-            return "ERROR"
-
-# --- Core Functions ---
+        question = (question or "").strip()
+        if not question:
+            raise ValueError("Digite uma pergunta para testar o agente.")
+
+        result = self.agent.run(question)
+        return str(result).replace("FINAL ANSWER", "").strip()
+
+
+def empty_results() -> pd.DataFrame:
+    return pd.DataFrame(columns=RESULT_COLUMNS)
+
+
+def test_agent(question: str):
+    """Executa uma única pergunta sem buscar nem enviar a avaliação oficial."""
+    try:
+        answer = BasicAgent()(question)
+        return "Teste concluído. Nenhum resultado foi enviado.", answer
+    except Exception as exc:
+        return f"Erro no teste: {exc}", ""
+
 
 def run_agent_only(profile: gr.OAuthProfile | None):
-    """
-    Busca as perguntas, roda o agente para gerar as respostas e retorna 
-    os dados para visualização e o payload para o estado do Gradio.
-    """
+    """Busca as perguntas e gera uma tabela editável, sem enviar respostas."""
     if not profile:
-        return "Please Login to Hugging Face first.", pd.DataFrame(), []
-
-    api_url = DEFAULT_API_URL
-    questions_url = f"{api_url}/questions"
+        return "Faça login no Hugging Face primeiro.", empty_results()
 
     try:
         agent = BasicAgent()
-    except Exception as e:
-        return f"Error initializing agent: {e}", pd.DataFrame(), []
+    except Exception as exc:
+        return f"Erro ao inicializar o agente: {exc}", empty_results()
 
-    print(f"Fetching questions from: {questions_url}")
+    questions_url = f"{DEFAULT_API_URL}/questions"
     try:
-        response = requests.get(questions_url, timeout=15)
+        response = requests.get(questions_url, timeout=30)
         response.raise_for_status()
-        questions_data = response.json()
-        if not questions_data:
-             return "Fetched questions list is empty.", pd.DataFrame(), []
-    except Exception as e:
-        return f"Error fetching questions: {e}", pd.DataFrame(), []
-
-    results_log = []
-    answers_payload = []
-    
-    print(f"Running agent on {len(questions_data)} questions...")
-    for item in questions_data:
+        questions = response.json()
+        if not questions:
+            return "A API retornou uma lista de perguntas vazia.", empty_results()
+    except Exception as exc:
+        return f"Erro ao buscar perguntas: {exc}", empty_results()
+
+    results = []
+    for item in questions:
         task_id = item.get("task_id")
-        question_text = item.get("question")
-        
-        if not task_id or question_text is None:
+        question = item.get("question")
+        if not task_id or question is None:
             continue
-            
+
         try:
-            submitted_answer = agent(question_text)
-            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}"})
-
-    if not answers_payload:
-        return "Agent did not produce any answers.", pd.DataFrame(results_log), []
-
-    status_update = f"Agent finished! Processed {len(answers_payload)} questions. Please review the table below before submitting."
-    results_df = pd.DataFrame(results_log)
-    
-    # Retorna o status, a tabela visível e o payload invisível (para o gr.State)
-    return status_update, results_df, answers_payload
-
-
-def submit_to_leaderboard(profile: gr.OAuthProfile | None, answers_payload: list):
-    """
-    Pega as respostas validadas no estado do Gradio e envia para a API.
-    """
+            answer = agent(question)
+        except Exception as exc:
+            answer = f"ERROR: {exc}"
+
+        results.append(
+            {
+                "Task ID": task_id,
+                "Question": question,
+                "Submitted Answer": answer,
+            }
+        )
+
+    if not results:
+        return "O agente não produziu respostas.", empty_results()
+
+    status = (
+        f"Execução concluída: {len(results)} respostas geradas. "
+        "Revise e, se necessário, edite a coluna 'Submitted Answer'. "
+        "Nada foi enviado ainda."
+    )
+    return status, pd.DataFrame(results, columns=RESULT_COLUMNS)
+
+
+def normalize_results(results_table) -> list[dict]:
+    """Converte a tabela revisada no payload exigido pela API."""
+    if results_table is None:
+        return []
+
+    if isinstance(results_table, pd.DataFrame):
+        dataframe = results_table.copy()
+    else:
+        dataframe = pd.DataFrame(results_table, columns=RESULT_COLUMNS)
+
+    if dataframe.empty:
+        return []
+
+    missing = set(RESULT_COLUMNS) - set(dataframe.columns)
+    if missing:
+        raise ValueError(
+            "A tabela de revisão não contém as colunas esperadas: "
+            + ", ".join(sorted(missing))
+        )
+
+    answers = []
+    for _, row in dataframe.iterrows():
+        task_id = str(row["Task ID"]).strip()
+        answer = str(row["Submitted Answer"]).strip()
+        if not task_id or task_id.lower() == "nan":
+            continue
+        if not answer or answer.lower() == "nan":
+            raise ValueError(f"A tarefa {task_id} está sem resposta.")
+        if answer.startswith("ERROR:"):
+            raise ValueError(
+                f"A tarefa {task_id} ainda contém um erro. "
+                "Corrija a resposta antes de enviar."
+            )
+        answers.append({"task_id": task_id, "submitted_answer": answer})
+
+    return answers
+
+
+def submit_to_leaderboard(
+    profile: gr.OAuthProfile | None, results_table
+):
+    """Envia exatamente os valores atualmente visíveis na tabela revisada."""
     if not profile:
-        return "Please Login to Hugging Face first."
-    
-    if not answers_payload or len(answers_payload) == 0:
-        return "No answers to submit. Please run the agent first."
-
-    username = profile.username
-    space_id = os.getenv("SPACE_ID", "local-environment")
-    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
-    submit_url = f"{DEFAULT_API_URL}/submit"
-
-    submission_data = {
-        "username": username.strip(), 
-        "agent_code": agent_code, 
-        "answers": answers_payload
+        return "Faça login no Hugging Face primeiro."
+
+    try:
+        answers = normalize_results(results_table)
+    except Exception as exc:
+        return f"Envio bloqueado: {exc}"
+
+    if not answers:
+        return "Não há respostas para enviar. Execute a avaliação primeiro."
+
+    space_id = os.getenv("SPACE_ID")
+    if not space_id:
+        return (
+            "Envio bloqueado: a variável SPACE_ID não foi encontrada. "
+            "Publique/execute este app em um Hugging Face Space."
+        )
+
+    submission = {
+        "username": profile.username.strip(),
+        "agent_code": f"https://huggingface.co/spaces/{space_id}/tree/main",
+        "answers": answers,
     }
-    
-    print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
+
     try:
-        response = requests.post(submit_url, json=submission_data, timeout=60)
+        response = requests.post(
+            f"{DEFAULT_API_URL}/submit",
+            json=submission,
+            timeout=90,
+        )
         response.raise_for_status()
-        result_data = response.json()
-        
-        final_status = (
-            f"✅ Submission Successful!\n"
-            f"User: {result_data.get('username')}\n"
-            f"Overall Score: {result_data.get('score', 'N/A')}% "
-            f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
-            f"Message: {result_data.get('message', 'No message received.')}"
+        result = response.json()
+        return (
+            "Envio realizado com sucesso!\n"
+            f"Usuário: {result.get('username')}\n"
+            f"Pontuação: {result.get('score', 'N/A')}% "
+            f"({result.get('correct_count', '?')}/"
+            f"{result.get('total_attempted', '?')} corretas)\n"
+            f"Mensagem: {result.get('message', 'Sem mensagem.')}"
         )
-        return final_status
-    except requests.exceptions.RequestException as e:
-        return f"❌ Submission Failed: {e}"
-    except Exception as e:
-        return f"❌ An unexpected error occurred: {e}"
+    except requests.exceptions.RequestException as exc:
+        detail = ""
+        if exc.response is not None:
+            detail = f" Resposta da API: {exc.response.text[:500]}"
+        return f"Falha no envio: {exc}.{detail}"
+    except Exception as exc:
+        return f"Erro inesperado no envio: {exc}"
 
 
-# --- Build Gradio Interface ---
-with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# GAIA Agent Evaluation - Two-Step Runner")
+with gr.Blocks(theme=gr.themes.Soft(), title="GAIA Agent Evaluation") as demo:
+    gr.Markdown("# GAIA Agent Evaluation")
     gr.Markdown(
-        """
-        **Instruções de Validação:**
-        1. Faça o Login no Hugging Face.
-        2. Clique em **'1. Run Agent & Preview'**. O agente vai processar as questões e a tabela será preenchida. (Isso pode demorar vários minutos).
-        3. Valide as respostas na tabela. Se o formato estiver correto (Exact Match), clique em **'2. Submit to Leaderboard'** para enviar sua pontuação oficial.
-        """
+        "Teste o agente isoladamente, gere as respostas oficiais para revisão "
+        "e só então faça o envio final."
     )
-
     gr.LoginButton()
 
-    with gr.Row():
-        btn_run = gr.Button("1. Run Agent & Preview Answers", variant="secondary")
-        btn_submit = gr.Button("2. Submit to Leaderboard (Final)", variant="primary")
+    with gr.Tabs():
+        with gr.Tab("1. Testar agente"):
+            gr.Markdown(
+                "Use uma pergunta livre para verificar o modelo e a busca. "
+                "Este teste não acessa nem envia a avaliação oficial."
+            )
+            test_question = gr.Textbox(
+                label="Pergunta de teste",
+                lines=5,
+                placeholder="Digite uma pergunta para o agente...",
+            )
+            test_button = gr.Button("Executar teste", variant="secondary")
+            test_status = gr.Textbox(label="Status", interactive=False)
+            test_answer = gr.Textbox(
+                label="Resposta do agente", lines=5, interactive=False
+            )
+            test_button.click(
+                fn=test_agent,
+                inputs=[test_question],
+                outputs=[test_status, test_answer],
+            )
 
-    status_output = gr.Textbox(label="System Status", lines=3, interactive=False)
-    results_table = gr.DataFrame(label="Agent Answers Preview", wrap=True)
-    
-    # Variável de estado invisível para armazenar o payload JSON entre os cliques dos botões
-    stored_answers = gr.State([])
+        with gr.Tab("2. Executar e revisar"):
+            gr.Markdown(
+                "Gere as respostas das 20 questões. Você pode editar a coluna "
+                "'Submitted Answer' antes do envio."
+            )
+            run_button = gr.Button(
+                "Executar avaliação sem enviar", variant="secondary"
+            )
+            run_status = gr.Textbox(label="Status", lines=4, interactive=False)
+            results_table = gr.DataFrame(
+                headers=RESULT_COLUMNS,
+                datatype=["str", "str", "str"],
+                value=empty_results(),
+                label="Respostas para revisão",
+                wrap=True,
+                interactive=True,
+            )
+            run_button.click(
+                fn=run_agent_only,
+                outputs=[run_status, results_table],
+            )
 
-    # Evento do Botão 1: Roda o agente, atualiza a interface e salva no estado
-    btn_run.click(
-        fn=run_agent_only,
-        outputs=[status_output, results_table, stored_answers]
-    )
+        with gr.Tab("3. Enviar resultado"):
+            gr.Markdown(
+                "O botão abaixo envia os valores atuais da tabela da aba "
+                "anterior. Confira todas as respostas antes de continuar."
+            )
+            submit_button = gr.Button(
+                "Enviar respostas revisadas ao leaderboard",
+                variant="primary",
+            )
+            submit_status = gr.Textbox(
+                label="Resultado do envio", lines=6, interactive=False
+            )
+            submit_button.click(
+                fn=submit_to_leaderboard,
+                inputs=[results_table],
+                outputs=[submit_status],
+            )
 
-    # Evento do Botão 2: Lê o estado e envia para a API
-    btn_submit.click(
-        fn=submit_to_leaderboard,
-        inputs=[stored_answers],
-        outputs=[status_output]
-    )
 
 if __name__ == "__main__":
-    print("\n" + "-"*30 + " App Starting " + "-"*30)
-    demo.launch(debug=True, share=False)
+    demo.launch(debug=True, share=False)