Update app.py

app.py

@@ -3,33 +3,102 @@ import gradio as gr
 import requests
 import pandas as pd
 import re
-import tempfile
 import pytesseract
-from PIL import Image
-from typing import Dict, List, Optional, TypedDict, Annotated
-from langgraph.graph import StateGraph, END
-from langgraph.checkpoint.memory import MemorySaver
-from langchain_core.messages import HumanMessage, SystemMessage, AnyMessage
-from langchain_openai import ChatOpenAI
-from langgraph.prebuilt import ToolNode, tools_condition
-from langchain_community.tools.tavily_search import TavilySearchResults
-from youtube_transcript_api import YouTubeTranscriptApi
-from langchain_core.tools import tool
 import yt_dlp
 import cv2
 import numpy as np
 import speech_recognition as sr
+from PIL import Image
+from typing import List, Dict
+from langchain_core.tools import tool
+from langchain_openai import ChatOpenAI
+from langchain_community.tools.tavily_search import TavilySearchResults
+from youtube_transcript_api import YouTubeTranscriptApi
+from langgraph.graph import StateGraph, END
+from langgraph.checkpoint.memory import MemorySaver
+from langchain_core.messages import HumanMessage, SystemMessage
 
-# ================ Configuración Global ================
+# ================ CONSTANTES ================
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
-USERNAME = "Csuarezg"  # Cambiar por tu usuario
+SYSTEM_PROMPT = SYSTEM_PROMPT = """You are a precision research assistant for the GAIA benchmark. Your mission is EXTREME ACCURACY.
+CRITICAL ANSWER FORMAT RULES:
+# - ALWAYS end with: FINAL ANSWER: [answer]
+# - READ THE QUESTION CAREFULLY - answer EXACTLY what is asked for, nothing more, nothing less
+SPECIFIC FORMATTING BY QUESTION TYPE:
+# - Numbers: ONLY the number, no units, no text
+# Example: "FINAL ANSWER: 2" NOT "FINAL ANSWER: 2 albums"
+# - First name only: ONLY the first name
+# Example: If person is "John Smith" → "FINAL ANSWER: John"
+# - Country codes, IOC codes, abbreviations, symbols: ONLY the code/abbreviation, no country name or brackets
+# Example: If asked for IOC country code → "FINAL ANSWER: PHI" NOT "FINAL ANSWER: PHILIPPINES [PHI]"
+# - When asked for a specific type of identifier (code, abbreviation, symbol):
+#   Give ONLY that identifier, strip all explanatory text, brackets, or full names
+# - Lists/Sets: Exactly as requested format
+# Example: "FINAL ANSWER: a, b, d, e" (comma-separated, alphabetical order)
+CRITICAL TOOL SELECTION:
+# - Wikipedia questions → wikipedia_tool ONLY
+# - File questions → file_analyzer_tool FIRST to inspect contents, then reason based on structure
+# - Current events → web_search_tool ONLY
+# - Mathematical analysis/calculations → wolfram_alpha_tool or python_repl_tool ONLY
+# - Tables, matrices, systematic checking → python_repl_tool ONLY
+FOR MATHEMATICAL PROBLEMS:
+# ALWAYS use python_repl_tool when:
+# - Analyzing mathematical tables or matrices
+# - Checking properties like commutativity, associativity
+# - Systematic verification of mathematical statements
+# - Complex calculations that need precision
+# - ANY problem involving tables, sets, or systematic checking
+MATHEMATICAL ANALYSIS PROCESS:
+# 1. Use python_repl_tool to parse data systematically
+# 2. Write code to check ALL cases (don't rely on manual inspection)
+# 3. Collect results programmatically
+# 4. Verify your logic with multiple approaches
+# 5. Format answer exactly as requested
+# Example for commutativity checking:
+# - Parse the operation table into a data structure
+# - Check ALL pairs (x,y) to see if x*y = y*x
+# - Collect ALL elements involved in ANY counter-example
+# - Return in requested format (e.g., comma-separated, alphabetical)
+FILE HANDLING:
+# - You HAVE the ability to read and analyze uploaded files
+# - ALWAYS use file_analyzer_tool when questions mention files
+# - The tool automatically finds and analyzes Excel, CSV, images, and audio files
+# - For Excel/CSV: Returns columns, data types, sample rows, and numeric totals
+# - NEVER say "I can't access files" - you CAN access them via file_analyzer_tool
+# - Example: "The attached Excel file..." → Use file_analyzer_tool immediately
+SPECIAL CASES TO HANDLE:
+# - If the question appears reversed or encoded, decode it first.
+# - If the question includes an instruction (e.g., "write the opposite of..."), follow the instruction precisely.
+# - DO NOT repeat or paraphrase the question in your answer.
+# - NEVER answer with the full sentence unless explicitly asked to.
+# - If the decoded question asks for a word, give ONLY the word, in the required format.
+REASONING PROCESS:
+# 1. Carefully read what the question is asking for
+# 2. Identify if it needs systematic/mathematical analysis
+# 3. Use appropriate tool (python_repl_tool for math problems)
+# 4. Extract ONLY the specific part requested
+# 5. Format according to the rules above
+# 6. For file questions:
+# a. First use file_analyzer_tool to inspect column names, types, and sample data
+# b. Identify relevant columns based on the question
+# c. Reason using the data (e.g., by counting, filtering, or identifying patterns)
+# d. Only use python_repl_tool if additional computation is necessary
+# 7. If the Wikipedia tool is used but fails to provide an answer (no relevant entry or content), automatically attempt a web search using the same query or a refined version of it
+"""
+USERNAME = "Csuarezg"
 AGENT_CODE = "gaia_agent_v1"
-SYSTEM_PROMPT = """[Insertar tu system prompt completo aquí]"""
 
-# ================ Definición de Herramientas con @tool ================
+# ================ HERRAMIENTAS ================
 @tool
 def wikipedia_tool(query: str) -> str:
-    """Busca en Wikipedia información enciclopédica."""
+    """Busca información enciclopédica en Wikipedia. Útil para datos históricos, biografías y conceptos científicos.
+    
+    Args:
+        query: Término de búsqueda específico (ej. 'Teoría de la relatividad')
+        
+    Returns:
+        Resumen conciso del tema en 3 oraciones.
+    """
     try:
         import wikipedia
         wikipedia.set_lang("en")
@@ -39,17 +108,31 @@ def wikipedia_tool(query: str) -> str:
 
 @tool
 def youtube_transcript_tool(url: str) -> str:
-    """Obtiene el transcript de un video de YouTube."""
+    """Obtiene el transcript de videos de YouTube. Útil para analizar diálogos o contenido hablado.
+    
+    Args:
+        url: Enlace completo del video (ej. 'https://youtu.be/VIDEO_ID')
+        
+    Returns:
+        Primera parte del transcript (primeros 30 segundos).
+    """
     try:
         video_id = re.findall(r'(?:v=|\/)([0-9A-Za-z_-]{11})', url)[0]
         transcript = YouTubeTranscriptApi.get_transcript(video_id)
         return " ".join([entry['text'] for entry in transcript[:5]])
     except Exception as e:
-        return f"Error de transcript: {str(e)}"
+        return f"Error transcript: {str(e)}"
 
 @tool
 def file_analyzer_tool(file_path: str) -> str:
-    """Analiza archivos de imagen y texto."""
+    """Analiza archivos (imágenes, audio) usando OCR y visión por computadora.
+    
+    Args:
+        file_path: Ruta al archivo en el sistema
+        
+    Returns:
+        Texto extraído o análisis de contenido multimedia.
+    """
     try:
         if file_path.endswith(('.png', '.jpg', '.jpeg')):
             img = Image.open(file_path)
@@ -57,110 +140,135 @@ def file_analyzer_tool(file_path: str) -> str:
             return f"Texto detectado: {text[:500]}..." if text else "Sin texto"
         return "Formato no soportado"
     except Exception as e:
-        return f"Error análisis de archivo: {str(e)}"
+        return f"Error análisis archivo: {str(e)}"
 
 @tool
 def web_search_tool(query: str) -> str:
-    """Realiza búsquedas web en tiempo real."""
+    """Realiza búsquedas web en tiempo real. Útil para información actualizada.
+    
+    Args:
+        query: Término de búsqueda con contexto
+        
+    Returns:
+        3 resultados relevantes con fuentes.
+    """
     try:
         tavily = TavilySearchResults(api_key=os.getenv("TAVILY_API_KEY"), max_results=3)
         results = tavily.invoke(query)
         return "\n".join([f"{res['title']}: {res['content']}" for res in results])
     except Exception as e:
-        return f"Error de búsqueda: {str(e)}"
+        return f"Error búsqueda: {str(e)}"
 
-# ================ Clase del Agente ================
+# ================ AGENTE PRINCIPAL ================
 class GaiaAgent:
     def __init__(self):
-        self.tools = [
-            wikipedia_tool,
-            youtube_transcript_tool,
-            file_analyzer_tool,
-            web_search_tool
-        ]
-        self.agent = self._create_agent()
-    
-    def _create_agent(self):
-        llm = ChatOpenAI(model="gpt-4-turbo", temperature=0)
-        model = llm.bind_tools(self.tools)
+        self.tools = [wikipedia_tool, youtube_transcript_tool, file_analyzer_tool, web_search_tool]
+        self.llm = ChatOpenAI(model="gpt-4-turbo", temperature=0)
+        self.workflow = self._build_workflow()
+        self.recognizer = sr.Recognizer()
+
+    def _build_workflow(self):
+        workflow = StateGraph(AgentState)
         
         def agent_node(state):
-            messages = state['messages']
-            if not any(isinstance(m, SystemMessage) for m in messages):
-                messages = [SystemMessage(content=SYSTEM_PROMPT)] + messages
-            response = model.invoke(messages)
+            messages = [SystemMessage(content=SYSTEM_PROMPT)] + state['messages']
+            response = self.llm.bind_tools(self.tools).invoke(messages)
             return {"messages": [response]}
         
-        tool_node = ToolNode(self.tools)
-        
-        workflow = StateGraph(AgentState)
         workflow.add_node("agent", agent_node)
-        workflow.add_node("tools", tool_node)
+        workflow.add_node("tools", ToolNode(self.tools))
+        
         workflow.set_entry_point("agent")
         workflow.add_conditional_edges(
             "agent",
-            lambda x: "tools" if x["messages"][-1].tool_calls else END,
+            lambda x: "tools" if x["messages"][-1].tool_calls else END
         )
         workflow.add_edge("tools", "agent")
         
         return workflow.compile()
-    
-    def process_question(self, question: str) -> str:
+
+    def __call__(self, question: str) -> str:
         try:
-            response = self.agent.invoke({"messages": [HumanMessage(content=question)]})
-            last_message = response["messages"][-1].content
-            return self._extract_final_answer(last_message)
+            response = self.workflow.invoke(
+                {"messages": [HumanMessage(content=question)]},
+                {"configurable": {"thread_id": "main_thread"}}
+            )
+            return self._extract_final_answer(response['messages'][-1].content)
         except Exception as e:
-            return f"Error del agente: {str(e)}"
-    
+            return f"Error: {str(e)}"
+
     def _extract_final_answer(self, text: str) -> str:
         match = re.search(r"FINAL ANSWER:\s*(.*)", text, re.IGNORECASE)
         return match.group(1).strip() if match else text
 
-# ================ Integración con Gradio ================
-def run_evaluation(profile: gr.OAuthProfile | None):
+# ================ LÓGICA DE EJECUCIÓN ================
+def run_and_submit_all(profile: gr.OAuthProfile | None):
+    space_id = os.getenv("SPACE_ID")
+    
     if not profile:
         return "Por favor inicia sesión primero", None
-    
+
     try:
         agent = GaiaAgent()
-        response = requests.get(f"{DEFAULT_API_URL}/questions")
-        questions = response.json()
+        questions_url = f"{DEFAULT_API_URL}/questions"
+        response = requests.get(questions_url, timeout=15)
+        response.raise_for_status()
+        questions_data = response.json()
         
         answers = []
-        results = []
-        for q in questions:
-            answer = agent.process_question(q["question"])
-            answers.append({"task_id": q["task_id"], "submitted_answer": answer})
-            results.append({"Pregunta": q["task_id"], "Respuesta": answer})
-        
-        submission = {
+        results_log = []
+        for item in questions_data:
+            task_id = item.get("task_id")
+            question_text = item.get("question")
+            if not task_id or not question_text:
+                continue
+            try:
+                answer = agent(question_text)
+                answers.append({"task_id": task_id, "submitted_answer": answer})
+                results_log.append({"Task ID": task_id, "Question": question_text, "Answer": answer})
+            except Exception as e:
+                results_log.append({"Task ID": task_id, "Question": question_text, "Answer": f"Error: {str(e)}"})
+
+        submission_data = {
             "username": USERNAME,
             "agent_code": AGENT_CODE,
             "answers": answers
         }
         
-        submit_response = requests.post(f"{DEFAULT_API_URL}/submit", json=submission)
+        submit_response = requests.post(f"{DEFAULT_API_URL}/submit", json=submission_data, timeout=60)
         submit_response.raise_for_status()
         
-        return f"Éxito! Puntaje: {submit_response.json().get('score', 0)}", pd.DataFrame(results)
+        result = submit_response.json()
+        status = (
+            f"¡Envío exitoso!\n"
+            f"Usuario: {result.get('username', '')}\n"
+            f"Puntaje: {result.get('score', 0)}%\n"
+            f"Mensaje: {result.get('message', '')}"
+        )
+        return status, pd.DataFrame(results_log)
     
     except Exception as e:
         return f"Error crítico: {str(e)}", pd.DataFrame()
 
-# ================ Interfaz de Usuario ================
-with gr.Blocks() as interface:
+# ================ INTERFAZ GRADIO ================
+with gr.Blocks() as demo:
     gr.Markdown("# GAIA Agent - Evaluación Completa")
-    login_btn = gr.LoginButton()
-    run_btn = gr.Button("Ejecutar Evaluación Completa")
-    status_output = gr.Textbox(label="Estado de la Ejecución", interactive=False)
-    results_table = gr.DataFrame(label="Resultados Detallados")
+    gr.Markdown("""
+    **Instrucciones:**
+    1. Inicia sesión con tu cuenta de Hugging Face
+    2. Haz clic en 'Ejecutar Evaluación'
+    3. Espera los resultados (puede tomar varios minutos)
+    """)
+    
+    gr.LoginButton()
+    run_btn = gr.Button("Ejecutar Evaluación", variant="primary")
+    status_output = gr.Textbox(label="Estado", interactive=False)
+    results_table = gr.DataFrame(label="Resultados Detallados", wrap=True)
     
     run_btn.click(
-        fn=run_evaluation,
-        inputs=[],
+        fn=run_and_submit_all,
         outputs=[status_output, results_table]
     )
 
 if __name__ == "__main__":
-    interface.launch(server_name="0.0.0.0", server_port=7860)
+    demo.launch(server_name="0.0.0.0", server_port=7860)