Update app.py
Browse files
app.py
CHANGED
|
@@ -3,8 +3,24 @@ import gradio as gr
|
|
| 3 |
import requests
|
| 4 |
import inspect
|
| 5 |
import pandas as pd
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
|
| 7 |
-
# (Keep Constants as is)
|
| 8 |
# --- Constants ---
|
| 9 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 10 |
|
|
@@ -13,182 +29,185 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
|
| 13 |
class BasicAgent:
|
| 14 |
def __init__(self):
|
| 15 |
print("BasicAgent initialized.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
def __call__(self, question: str) -> str:
|
| 17 |
-
print(f"
|
| 18 |
-
fixed_answer = "This is a default answer."
|
| 19 |
-
print(f"Agent returning fixed answer: {fixed_answer}")
|
| 20 |
-
return fixed_answer
|
| 21 |
-
|
| 22 |
-
def run_and_submit_all( profile: gr.OAuthProfile | None):
|
| 23 |
-
"""
|
| 24 |
-
Fetches all questions, runs the BasicAgent on them, submits all answers,
|
| 25 |
-
and displays the results.
|
| 26 |
-
"""
|
| 27 |
-
# --- Determine HF Space Runtime URL and Repo URL ---
|
| 28 |
-
space_id = os.getenv("SPACE_ID") # Get the SPACE_ID for sending link to the code
|
| 29 |
-
|
| 30 |
-
if profile:
|
| 31 |
-
username= f"{profile.username}"
|
| 32 |
-
print(f"User logged in: {username}")
|
| 33 |
-
else:
|
| 34 |
-
print("User not logged in.")
|
| 35 |
-
return "Please Login to Hugging Face with the button.", None
|
| 36 |
-
|
| 37 |
-
api_url = DEFAULT_API_URL
|
| 38 |
-
questions_url = f"{api_url}/questions"
|
| 39 |
-
submit_url = f"{api_url}/submit"
|
| 40 |
-
|
| 41 |
-
# 1. Instantiate Agent ( modify this part to create your agent)
|
| 42 |
-
try:
|
| 43 |
-
agent = BasicAgent()
|
| 44 |
-
except Exception as e:
|
| 45 |
-
print(f"Error instantiating agent: {e}")
|
| 46 |
-
return f"Error initializing agent: {e}", None
|
| 47 |
-
# In the case of an app running as a hugging Face space, this link points toward your codebase ( usefull for others so please keep it public)
|
| 48 |
-
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
|
| 49 |
-
print(agent_code)
|
| 50 |
-
|
| 51 |
-
# 2. Fetch Questions
|
| 52 |
-
print(f"Fetching questions from: {questions_url}")
|
| 53 |
-
try:
|
| 54 |
-
response = requests.get(questions_url, timeout=15)
|
| 55 |
-
response.raise_for_status()
|
| 56 |
-
questions_data = response.json()
|
| 57 |
-
if not questions_data:
|
| 58 |
-
print("Fetched questions list is empty.")
|
| 59 |
-
return "Fetched questions list is empty or invalid format.", None
|
| 60 |
-
print(f"Fetched {len(questions_data)} questions.")
|
| 61 |
-
except requests.exceptions.RequestException as e:
|
| 62 |
-
print(f"Error fetching questions: {e}")
|
| 63 |
-
return f"Error fetching questions: {e}", None
|
| 64 |
-
except requests.exceptions.JSONDecodeError as e:
|
| 65 |
-
print(f"Error decoding JSON response from questions endpoint: {e}")
|
| 66 |
-
print(f"Response text: {response.text[:500]}")
|
| 67 |
-
return f"Error decoding server response for questions: {e}", None
|
| 68 |
-
except Exception as e:
|
| 69 |
-
print(f"An unexpected error occurred fetching questions: {e}")
|
| 70 |
-
return f"An unexpected error occurred fetching questions: {e}", None
|
| 71 |
-
|
| 72 |
-
# 3. Run your Agent
|
| 73 |
-
results_log = []
|
| 74 |
-
answers_payload = []
|
| 75 |
-
print(f"Running agent on {len(questions_data)} questions...")
|
| 76 |
-
for item in questions_data:
|
| 77 |
-
task_id = item.get("task_id")
|
| 78 |
-
question_text = item.get("question")
|
| 79 |
-
if not task_id or question_text is None:
|
| 80 |
-
print(f"Skipping item with missing task_id or question: {item}")
|
| 81 |
-
continue
|
| 82 |
try:
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
except Exception as e:
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
print(status_update)
|
| 98 |
-
|
| 99 |
-
# 5. Submit
|
| 100 |
-
print(f"Submitting {len(answers_payload)} answers to: {submit_url}")
|
| 101 |
-
try:
|
| 102 |
-
response = requests.post(submit_url, json=submission_data, timeout=60)
|
| 103 |
-
response.raise_for_status()
|
| 104 |
-
result_data = response.json()
|
| 105 |
-
final_status = (
|
| 106 |
-
f"Submission Successful!\n"
|
| 107 |
-
f"User: {result_data.get('username')}\n"
|
| 108 |
-
f"Overall Score: {result_data.get('score', 'N/A')}% "
|
| 109 |
-
f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
|
| 110 |
-
f"Message: {result_data.get('message', 'No message received.')}"
|
| 111 |
-
)
|
| 112 |
-
print("Submission successful.")
|
| 113 |
-
results_df = pd.DataFrame(results_log)
|
| 114 |
-
return final_status, results_df
|
| 115 |
-
except requests.exceptions.HTTPError as e:
|
| 116 |
-
error_detail = f"Server responded with status {e.response.status_code}."
|
| 117 |
-
try:
|
| 118 |
-
error_json = e.response.json()
|
| 119 |
-
error_detail += f" Detail: {error_json.get('detail', e.response.text)}"
|
| 120 |
-
except requests.exceptions.JSONDecodeError:
|
| 121 |
-
error_detail += f" Response: {e.response.text[:500]}"
|
| 122 |
-
status_message = f"Submission Failed: {error_detail}"
|
| 123 |
-
print(status_message)
|
| 124 |
-
results_df = pd.DataFrame(results_log)
|
| 125 |
-
return status_message, results_df
|
| 126 |
-
except requests.exceptions.Timeout:
|
| 127 |
-
status_message = "Submission Failed: The request timed out."
|
| 128 |
-
print(status_message)
|
| 129 |
-
results_df = pd.DataFrame(results_log)
|
| 130 |
-
return status_message, results_df
|
| 131 |
-
except requests.exceptions.RequestException as e:
|
| 132 |
-
status_message = f"Submission Failed: Network error - {e}"
|
| 133 |
-
print(status_message)
|
| 134 |
-
results_df = pd.DataFrame(results_log)
|
| 135 |
-
return status_message, results_df
|
| 136 |
-
except Exception as e:
|
| 137 |
-
status_message = f"An unexpected error occurred during submission: {e}"
|
| 138 |
-
print(status_message)
|
| 139 |
-
results_df = pd.DataFrame(results_log)
|
| 140 |
-
return status_message, results_df
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
# --- Build Gradio Interface using Blocks ---
|
| 144 |
-
with gr.Blocks() as demo:
|
| 145 |
-
gr.Markdown("# Basic Agent Evaluation Runner")
|
| 146 |
-
gr.Markdown(
|
| 147 |
-
"""
|
| 148 |
-
**Instructions:**
|
| 149 |
-
1. Please clone this space, then modify the code to define your agent's logic, the tools, the necessary packages, etc ...
|
| 150 |
-
2. Log in to your Hugging Face account using the button below. This uses your HF username for submission.
|
| 151 |
-
3. Click 'Run Evaluation & Submit All Answers' to fetch questions, run your agent, submit answers, and see the score.
|
| 152 |
-
---
|
| 153 |
-
**Disclaimers:**
|
| 154 |
-
Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
|
| 155 |
-
This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
|
| 156 |
-
"""
|
| 157 |
-
)
|
| 158 |
-
|
| 159 |
-
gr.LoginButton()
|
| 160 |
-
|
| 161 |
-
run_button = gr.Button("Run Evaluation & Submit All Answers")
|
| 162 |
-
|
| 163 |
-
status_output = gr.Textbox(label="Run Status / Submission Result", lines=5, interactive=False)
|
| 164 |
-
# Removed max_rows=10 from DataFrame constructor
|
| 165 |
-
results_table = gr.DataFrame(label="Questions and Agent Answers", wrap=True)
|
| 166 |
-
|
| 167 |
-
run_button.click(
|
| 168 |
-
fn=run_and_submit_all,
|
| 169 |
-
outputs=[status_output, results_table]
|
| 170 |
-
)
|
| 171 |
-
|
| 172 |
-
if __name__ == "__main__":
|
| 173 |
-
print("\n" + "-"*30 + " App Starting " + "-"*30)
|
| 174 |
-
# Check for SPACE_HOST and SPACE_ID at startup for information
|
| 175 |
-
space_host_startup = os.getenv("SPACE_HOST")
|
| 176 |
-
space_id_startup = os.getenv("SPACE_ID") # Get SPACE_ID at startup
|
| 177 |
-
|
| 178 |
-
if space_host_startup:
|
| 179 |
-
print(f"✅ SPACE_HOST found: {space_host_startup}")
|
| 180 |
-
print(f" Runtime URL should be: https://{space_host_startup}.hf.space")
|
| 181 |
-
else:
|
| 182 |
-
print("ℹ️ SPACE_HOST environment variable not found (running locally?).")
|
| 183 |
-
|
| 184 |
-
if space_id_startup: # Print repo URLs if SPACE_ID is found
|
| 185 |
-
print(f"✅ SPACE_ID found: {space_id_startup}")
|
| 186 |
-
print(f" Repo URL: https://huggingface.co/spaces/{space_id_startup}")
|
| 187 |
-
print(f" Repo Tree URL: https://huggingface.co/spaces/{space_id_startup}/tree/main")
|
| 188 |
-
else:
|
| 189 |
-
print("ℹ️ SPACE_ID environment variable not found (running locally?). Repo URL cannot be determined.")
|
| 190 |
-
|
| 191 |
-
print("-"*(60 + len(" App Starting ")) + "\n")
|
| 192 |
-
|
| 193 |
-
print("Launching Gradio Interface for Basic Agent Evaluation...")
|
| 194 |
-
demo.launch(debug=True, share=False)
|
|
|
|
| 3 |
import requests
|
| 4 |
import inspect
|
| 5 |
import pandas as pd
|
| 6 |
+
from typing import List, Dict
|
| 7 |
+
|
| 8 |
+
# --- Importar las librerías necesarias para el agente ---
|
| 9 |
+
# Para el LLM de Google
|
| 10 |
+
from langchain_google_genai import ChatGoogleGenerativeAI # Para modelos de chat como Gemini
|
| 11 |
+
# Para construir el agente
|
| 12 |
+
from langchain.agents import AgentExecutor, create_react_agent
|
| 13 |
+
from langchain import hub # Para jalar prompts estándar
|
| 14 |
+
from langchain.tools import Tool # Para envolver tus funciones como herramientas
|
| 15 |
+
from langchain_core.prompts import PromptTemplate # Para personalizar el prompt
|
| 16 |
+
# Para la herramienta de búsqueda web
|
| 17 |
+
from langchain_community.tools.tavily_research import TavilySearchResults
|
| 18 |
+
# Para manejar variables de entorno (opcional si solo usas secretos de HF)
|
| 19 |
+
from dotenv import load_dotenv
|
| 20 |
+
|
| 21 |
+
# Cargar variables de entorno si estás desarrollando localmente
|
| 22 |
+
# load_dotenv() # Descomenta si desarrollas localmente y tienes .env
|
| 23 |
|
|
|
|
| 24 |
# --- Constants ---
|
| 25 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 26 |
|
|
|
|
| 29 |
class BasicAgent:
|
| 30 |
def __init__(self):
|
| 31 |
print("BasicAgent initialized.")
|
| 32 |
+
|
| 33 |
+
# --- 1. Configuración de las Claves API ---
|
| 34 |
+
# Se leerán automáticamente de los secretos del Space de HF o de .env si load_dotenv() se usa.
|
| 35 |
+
google_api_key = os.getenv("GOOGLE_API_KEY")
|
| 36 |
+
tavily_api_key = os.getenv("TAVILY_API_KEY")
|
| 37 |
+
|
| 38 |
+
if not google_api_key:
|
| 39 |
+
raise ValueError("GOOGLE_API_KEY environment variable not set. Please add it to your Hugging Face Space secrets.")
|
| 40 |
+
if not tavily_api_key:
|
| 41 |
+
raise ValueError("TAVILY_API_KEY environment variable not set. Please add it to your Hugging Face Space secrets.")
|
| 42 |
+
|
| 43 |
+
# --- 2. Inicialización del LLM (Gemini Pro) ---
|
| 44 |
+
# Puedes probar con "gemini-pro" o "gemini-1.5-pro" si tienes acceso y tu Space lo soporta.
|
| 45 |
+
self.llm = ChatGoogleGenerativeAI(
|
| 46 |
+
model="gemini-pro",
|
| 47 |
+
google_api_key=google_api_key,
|
| 48 |
+
temperature=0.1, # Más bajo para respuestas más consistentes y precisas
|
| 49 |
+
max_tokens=2048 # Aumenta si necesitas respuestas más largas
|
| 50 |
+
)
|
| 51 |
+
print("LLM (Gemini Pro) initialized.")
|
| 52 |
+
|
| 53 |
+
# --- 3. Definición de Herramientas ---
|
| 54 |
+
# a) Tavily Search (MUY IMPORTANTE para GAIA)
|
| 55 |
+
self.tavily_search_tool = TavilySearchResults(api_key=tavily_api_key, max_results=5) # max_results limita los resultados
|
| 56 |
+
|
| 57 |
+
# b) Herramienta para descargar y leer archivos (adaptada de nuestra conversación anterior)
|
| 58 |
+
# Asegúrate de que DEFAULT_API_URL esté accesible en el scope de la clase
|
| 59 |
+
self.api_url = DEFAULT_API_URL
|
| 60 |
+
|
| 61 |
+
def download_and_read_task_file(task_id: str, filename: str) -> str:
|
| 62 |
+
"""
|
| 63 |
+
Downloads a file associated with a task_id from the API and returns its content.
|
| 64 |
+
Supports text, CSV, JSON, and PDF files.
|
| 65 |
+
Args:
|
| 66 |
+
task_id (str): The ID of the task associated with the file.
|
| 67 |
+
filename (str): The name of the file to download (e.g., "document.pdf", "data.csv").
|
| 68 |
+
Returns:
|
| 69 |
+
str: The content of the file as a string.
|
| 70 |
+
"""
|
| 71 |
+
file_url = f"{self.api_url}/files/{task_id}"
|
| 72 |
+
print(f"DEBUG: Attempting to download file from: {file_url} (filename: {filename})")
|
| 73 |
+
try:
|
| 74 |
+
response = requests.get(file_url, timeout=30, stream=True)
|
| 75 |
+
response.raise_for_status() # Raise an exception for bad status codes
|
| 76 |
+
|
| 77 |
+
temp_filepath = f"/tmp/{filename}" # Use /tmp for ephemeral storage in Space
|
| 78 |
+
with open(temp_filepath, 'wb') as f:
|
| 79 |
+
for chunk in response.iter_content(chunk_size=8192):
|
| 80 |
+
f.write(chunk)
|
| 81 |
+
|
| 82 |
+
content = ""
|
| 83 |
+
# Process the file based on its extension
|
| 84 |
+
if filename.lower().endswith(".pdf"):
|
| 85 |
+
try:
|
| 86 |
+
import pymupdf # type: ignore
|
| 87 |
+
doc = pymupdf.open(temp_filepath)
|
| 88 |
+
for page in doc:
|
| 89 |
+
content += page.get_text()
|
| 90 |
+
doc.close()
|
| 91 |
+
except ImportError:
|
| 92 |
+
content = "Error: pymupdf not installed for PDF reading."
|
| 93 |
+
except Exception as e:
|
| 94 |
+
content = f"Error reading PDF: {e}"
|
| 95 |
+
elif filename.lower().endswith((".csv", ".txt", ".json", ".log", ".md")):
|
| 96 |
+
with open(temp_filepath, 'r', encoding='utf-8') as f:
|
| 97 |
+
content = f.read()
|
| 98 |
+
elif filename.lower().endswith((".png", ".jpg", ".jpeg", ".gif")):
|
| 99 |
+
# Para imágenes, necesitarías OCR. Esto es más complejo.
|
| 100 |
+
# Por ahora, solo indicamos que se descargó.
|
| 101 |
+
content = f"Image file '{filename}' downloaded. OCR not implemented."
|
| 102 |
+
else:
|
| 103 |
+
content = f"Unsupported file type for {filename}. Cannot extract content."
|
| 104 |
+
|
| 105 |
+
os.remove(temp_filepath) # Clean up the temporary file
|
| 106 |
+
|
| 107 |
+
return f"Content of {filename}:\n{content[:1000]}..." # Return first 1000 chars for brevity
|
| 108 |
+
except requests.exceptions.RequestException as e:
|
| 109 |
+
return f"Error downloading file '{filename}' from {file_url}: {e}"
|
| 110 |
+
except Exception as e:
|
| 111 |
+
return f"An unexpected error occurred processing file '{filename}': {e}"
|
| 112 |
+
|
| 113 |
+
self.tools = [
|
| 114 |
+
Tool(
|
| 115 |
+
name="TavilySearch",
|
| 116 |
+
func=self.tavily_search_tool.run,
|
| 117 |
+
description="Útil para buscar información general, hechos, noticias y definiciones en internet. Siempre usa esta herramienta cuando necesites información externa.",
|
| 118 |
+
),
|
| 119 |
+
Tool(
|
| 120 |
+
name="DownloadAndReadTaskFile",
|
| 121 |
+
func=download_and_read_task_file,
|
| 122 |
+
description="Descarga y lee el contenido de un archivo asociado a una tarea específica. Necesita 'task_id' (string) y 'filename' (string, e.g., 'document.pdf', 'data.csv'). Útil cuando la pregunta menciona archivos.",
|
| 123 |
+
),
|
| 124 |
+
# Puedes añadir más herramientas aquí:
|
| 125 |
+
# - Python REPL si necesitas ejecutar código Python
|
| 126 |
+
# - Wikipedia si quieres una fuente de conocimiento enciclopédico
|
| 127 |
+
# - Herramienta de cálculo, etc.
|
| 128 |
+
]
|
| 129 |
+
print(f"Agent initialized with {len(self.tools)} tools.")
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
# --- 4. Construir el Agente ReAct ---
|
| 133 |
+
# El prompt estándar de ReAct para modelos de chat
|
| 134 |
+
# prompt = hub.pull("hwchase17/react-chat") # Puedes probar este también
|
| 135 |
+
|
| 136 |
+
# Customización del prompt para "EXACT MATCH"
|
| 137 |
+
custom_prompt_template = """Responde la siguiente pregunta de forma concisa y directa. Tu objetivo es proporcionar un "EXACT MATCH" con la respuesta correcta.
|
| 138 |
+
|
| 139 |
+
INSTRUCCIONES CLAVE:
|
| 140 |
+
- Si la pregunta requiere un número, solo devuelve el número (ej: 12345).
|
| 141 |
+
- Si requiere una fecha, solo la fecha en el formato solicitado o más común (ej: 2023-10-26 o October 26, 2023).
|
| 142 |
+
- Si requiere un nombre, solo el nombre (ej: París).
|
| 143 |
+
- Si requiere una lista, solo la lista de elementos separados por comas o líneas, sin numeración ni viñetas, a menos que se especifique lo contrario (ej: Manzanas, Peras, Uvas).
|
| 144 |
+
- NO incluyas ninguna explicación, introducción, despedida, o texto adicional.
|
| 145 |
+
- Utiliza las herramientas disponibles cuando sea necesario para encontrar la información.
|
| 146 |
+
- Piensa paso a paso y sé muy preciso en tu razonamiento.
|
| 147 |
+
- Si la pregunta menciona un archivo, utiliza la herramienta `DownloadAndReadTaskFile` con el `task_id` y el nombre del archivo.
|
| 148 |
+
|
| 149 |
+
Question: {input}
|
| 150 |
+
{agent_scratchpad}"""
|
| 151 |
+
|
| 152 |
+
self.prompt = PromptTemplate.from_template(custom_prompt_template)
|
| 153 |
+
|
| 154 |
+
# Para create_react_agent, necesitamos decirle qué herramientas están disponibles y sus nombres.
|
| 155 |
+
# Esto se hace automáticamente si usas hub.pull("hwchase17/react-chat-json") o similar
|
| 156 |
+
# Pero con un prompt customizado, lo hacemos explícito.
|
| 157 |
+
self.prompt = self.prompt.partial(
|
| 158 |
+
tools="\n".join([f"{tool.name}: {tool.description}" for tool in self.tools]),
|
| 159 |
+
tool_names=", ".join([tool.name for tool in self.tools]),
|
| 160 |
+
)
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
self.agent = create_react_agent(self.llm, self.tools, self.prompt)
|
| 164 |
+
# verbose=True es ESENCIAL para depurar y ver el razonamiento del agente.
|
| 165 |
+
# handle_parsing_errors=True para que el agente intente recuperarse de errores de formato.
|
| 166 |
+
self.agent_executor = AgentExecutor(
|
| 167 |
+
agent=self.agent,
|
| 168 |
+
tools=self.tools,
|
| 169 |
+
verbose=True,
|
| 170 |
+
handle_parsing_errors=True,
|
| 171 |
+
max_iterations=15, # Limita el número de pasos para evitar bucles infinitos
|
| 172 |
+
agent_kwargs={"handle_parsing_errors": True} # También en el agente
|
| 173 |
+
)
|
| 174 |
+
print("AgentExecutor initialized.")
|
| 175 |
+
|
| 176 |
def __call__(self, question: str) -> str:
|
| 177 |
+
print(f"\n--- Processing new question (first 100 chars): {question[:100]}...")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
try:
|
| 179 |
+
# Si la pregunta incluye "Task ID: [ID]", extrae el task_id.
|
| 180 |
+
# GAIA a menudo hace esto para indicar el archivo.
|
| 181 |
+
task_id_match = re.search(r"Task ID:\s*([\w-]+)", question)
|
| 182 |
+
current_task_id = task_id_match.group(1) if task_id_match else None
|
| 183 |
+
print(f"Detected Task ID for question: {current_task_id}")
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
# Invocar al agente. Aquí es donde el agente usa el LLM y las herramientas.
|
| 187 |
+
# Pasamos el task_id para que la herramienta de descarga de archivos pueda usarlo si es necesario.
|
| 188 |
+
# Algunos agentes pueden requerir un formato de input diferente, ReAct usa "input".
|
| 189 |
+
response = self.agent_executor.invoke({"input": question, "task_id": current_task_id})
|
| 190 |
+
# La respuesta final del agente React suele estar en 'output'
|
| 191 |
+
final_answer = response.get("output", "No answer generated by agent.")
|
| 192 |
+
|
| 193 |
+
# Post-procesamiento para asegurar EXACT MATCH y eliminar texto adicional
|
| 194 |
+
# Esto es CRÍTICO para el scoring.
|
| 195 |
+
final_answer = final_answer.strip()
|
| 196 |
+
# Elimina cualquier texto como "La respuesta es:", "The final answer is:", etc.
|
| 197 |
+
# Haz esto solo si detectas que el LLM aún añade verbosidad.
|
| 198 |
+
# if final_answer.lower().startswith(("the answer is", "la respuesta es", "final answer:")):
|
| 199 |
+
# final_answer = re.sub(r"^[Tt]he [Aa]nswer is: |^[Ll]a [Rr]espuesta es: |^[Ff]inal [Aa]nswer: ", "", final_answer, flags=re.IGNORECASE).strip()
|
| 200 |
+
|
| 201 |
+
print(f"Agent returning final answer: '{final_answer}'")
|
| 202 |
+
return final_answer
|
| 203 |
except Exception as e:
|
| 204 |
+
print(f"Error during agent execution: {e}")
|
| 205 |
+
# En caso de error, puedes devolver un mensaje de error o una cadena vacía.
|
| 206 |
+
# Para el scoring, una cadena vacía o un error puede contar como incorrecto.
|
| 207 |
+
return f"ERROR: {e}"
|
| 208 |
+
|
| 209 |
+
# El resto del código de `run_and_submit_all` y Gradio permanece igual.
|
| 210 |
+
# ... (código existente) ...
|
| 211 |
+
|
| 212 |
+
# Asegúrate de importar 're' al principio del archivo si lo usas para task_id_match
|
| 213 |
+
import re
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|