app.py

# -*- coding: utf-8 -*-
import os
import time
import requests
import pandas as pd
import gradio as gr
from dotenv import load_dotenv
from smolagents import CodeAgent, LiteLLMModel, tool

# URL da API responsável por fornecer as perguntas e receber o envio do benchmark.
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"

class BasicAgent:
    def __init__(self):
        # 1. Carrega variáveis de ambiente (útil para testes locais, ignorado no HF Spaces)
        load_dotenv()

        # 2. Configuração do Langfuse (Opcional - não vai quebrar se falhar)
        try:
            from langfuse import get_client
            from openinference.instrumentation.smolagents import SmolagentsInstrumentor

            langfuse_client = get_client()
            if langfuse_client.auth_check():
                print("📡 Langfuse autenticado com sucesso!")
                SmolagentsInstrumentor().instrument()
            else:
                print("⚠️ Langfuse ignorado (chaves ausentes).")
        except Exception:
            print("⚠️ Monitoramento do Langfuse desativado.")

        # Valida a presença da chave do Gemini (Obrigatório para o cérebro)
        gemini_key = os.getenv("GEMINI_API_KEY")
        if not gemini_key:
            print("❌ ERRO: A variável 'GEMINI_API_KEY' não foi encontrada nos Secrets.")

        # 3. Inicialização do modelo LLM usando LiteLLM (Corrigido para Gemini 2.0 Flash)
        self.model = LiteLLMModel(
            model_id="gemini/gemini-2.0-flash",
            api_key=gemini_key,
            num_retries=3 # Resiliência contra o erro 429
        )

        # 4. Ferramenta de busca Web
        @tool
        def busca_web(query: str) -> str:
            """Useful to search the web for up-to-date facts, Wikipedia articles, or general information.
            Args:
                query: The exact search query to look up on the internet.
            """
            try:
                headers = {'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) Chrome/120.0.0.0 Safari/537.36'}
                url = f"https://html.duckduckgo.com/html/?q={requests.utils.quote(query)}"
                res = requests.get(url, headers=headers, timeout=15)
                res.raise_for_status()

                from bs4 import BeautifulSoup
                soup = BeautifulSoup(res.text, 'html.parser')
                snippets = [span.get_text() for span in soup.find_all('span', class_='result__snippet')]
                
                if not snippets:
                    return "No clear results found on the web for this query."
                return "\n\n".join(snippets[:4])
            except Exception as e:
                return f"Search failed due to network error: {e}"

        # 5. Ferramenta de transcrição de áudio via OpenAI Whisper
        @tool
        def transcribe_audio(file_path: str) -> str:
            """Useful to transcribe any audio file (like MP3, WAV, M4A) into text.
            Always use this tool first when a question involves understanding audio.
            Args:
                file_path: The local path to the audio file (e.g., 'audio.mp3').
            """
            openai_key = os.getenv("OPENAI_API_KEY")
            if not openai_key:
                return "Error: OPENAI_API_KEY not found in secrets. Cannot transcribe audio."
            
            try:
                import openai
                client = openai.OpenAI(api_key=openai_key)
                with open(file_path, "rb") as audio_file:
                    transcript = client.audio.transcriptions.create(
                        model="whisper-1",
                        file=audio_file
                    )
                return f"Audio Transcription Content:\n{transcript.text}"
            except Exception as e:
                return f"Error transcribing audio: {e}."

        # 6. Instanciação do CodeAgent
        self.agent = CodeAgent(
            tools=[busca_web, transcribe_audio],
            model=self.model,
            add_base_tools=False,
            max_steps=10, # Adicionado limite de passos para evitar loops infinitos
            additional_authorized_imports=[
                "requests", "pydub", "wave", "openai",
                "PIL", "pdfplumber", "pypdf",
                "json", "csv", "openpyxl", "pandas",
                "os", "pathlib", "zipfile",
                "math", "datetime", "re", "itertools", "bs4"
            ]
        )

    def __call__(self, question: str) -> str:
        """Permite chamar o agente diretamente passando a pergunta."""
        print(f"Agent received question (first 50 chars): {question[:50]}...")

        # Prompt ajustado com regras rígidas para o GAIA
        prompt_ajustado = (
            f"TASK TO SOLVE: {question}\n\n"
            "EXECUTION RULES:\n"
            "1. You MUST solve this task step-by-step using Python code.\n"
            "2. Every single response you generate MUST strictly follow this exact grammar:\n"
            "Thoughts: <your reasoning here>\n"
            "<code>\n"
            "# your python code here using available tools\n"
            "</code>\n"
            "3. NEVER write conversational text or explanations outside of the 'Thoughts' or '<code>' sections.\n"
            "4. To finish the task and deliver the answer, you MUST call the `final_answer` tool inside a code block.\n"
            "5. CRITICAL FOR GAIA BENCHMARK (EXACT MATCH STRICT RULE):\n"
            "Inside the `final_answer()` tool, pass ONLY the raw string or number value matching the exact required format. Do NOT add labels or conversational prefixes.\n\n"
            "FEW-SHOT EXAMPLES OF EXPECTED FINAL ANSWERS:\n"
            "- Question: What was the actual enrollment count of the clinical trial on H. pylori in acne vulgaris patients from Jan-May 2018 as listed on the NIH website?\n"
            "  Correct Call: final_answer(90) or final_answer('90')\n\n"
            "- Question: If this whole pint is made up of ice cream, how many percent above or below the US federal standards for butterfat content is it when using the standards as reported by Wikipedia in 2020? Answer as + or - a number rounded to one decimal place.\n"
            "  Correct Call: final_answer('+4.6')\n\n"
            "- Question: In NASA's Astronomy Picture of the Day on 2006 January 21, two astronauts are visible... Give the last name of the astronaut, separated from the number of minutes by a semicolon.\n"
            "  Correct Call: final_answer('White; 5876')\n\n"
            "6. WEB REQUESTS: Always provide a User-Agent header when using `requests.get()` to avoid 403 Forbidden errors."
        )

        max_tentativas = 2 # Reduzido de 3 para 2 para evitar estourar cota de tempo atoa
        segundos_de_espera = 15

        for tentativa in range(max_tentativas):
            try:
                resposta_final = self.agent.run(prompt_ajustado)
                texto_resposta = str(resposta_final).strip()

                # Sanitização
                prefixos_para_remover = [
                    "final answer:", "final answer",
                    "the final answer is:", "the final answer is",
                    "answer:", "the answer is:"
                ]
                texto_lower = texto_resposta.lower()
                for prefixo in prefixos_para_remover:
                    if texto_lower.startswith(prefixo):
                        texto_resposta = texto_resposta[len(prefixo):].strip()
                        texto_lower = texto_resposta.lower()

                texto_resposta = texto_resposta.strip(" \t\n\r:.\"'")
                return texto_resposta

            except Exception as e:
                print(f"⚠️ Falha na tentativa {tentativa + 1}/{max_tentativas}: {e}")
                if tentativa < max_tentativas - 1:
                    time.sleep(segundos_de_espera)
                else:
                    return f"Erro definitivo da API: {e}"

def run_and_submit_all(profile: gr.OAuthProfile | None):
    # Resgata automaticamente os dados do Space atual
    username = os.getenv("SPACE_AUTHOR_NAME", "marantmir")
    space_id = os.getenv("SPACE_ID", f"{username}/Final_Assignment_Template")
    agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"

    questions_url = f"{DEFAULT_API_URL}/questions"
    submit_url = f"{DEFAULT_API_URL}/submit"

    try:
        agent = BasicAgent()
    except Exception as e:
        return f"Error initializing agent: {e}", None

    try:
        response = requests.get(questions_url, timeout=15)
        response.raise_for_status()
        questions_data = response.json()
        if not questions_data:
             return "Fetched questions list is empty.", None
    except Exception as e:
        return f"Error fetching questions: {e}", None

    results_log = []
    answers_payload = []
    for item in questions_data:
        task_id = item.get("task_id")
        question_text = item.get("question")
        if not task_id or question_text 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)

    submission_data = {"username": username, "agent_code": agent_code, "answers": answers_payload}

    try:
        response = requests.post(submit_url, json=submission_data, timeout=60)
        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')}%\n"
            f"Message: {result_data.get('message', '')}"
        )
        return final_status, pd.DataFrame(results_log)
    except Exception as e:
        return f"Submission Failed: {e}", pd.DataFrame(results_log)

# Interface Gradio
with gr.Blocks() as demo:
    gr.Markdown("# GAIA Benchmark - SmolAgents Runner")
    run_button = gr.Button("Run Evaluation & Submit All Answers")
    status_output = gr.Textbox(label="Run Status", interactive=False)
    results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
    run_button.click(fn=run_and_submit_all, outputs=[status_output, results_table])

if __name__ == "__main__":
    demo.launch(debug=True, share=False)