Spaces:
Runtime error
Runtime error
fff
Browse files- my_tools.py +179 -120
my_tools.py
CHANGED
|
@@ -1,3 +1,5 @@
|
|
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
import math
|
| 3 |
import time
|
|
@@ -5,7 +7,7 @@ import asyncio
|
|
| 5 |
import subprocess
|
| 6 |
import requests
|
| 7 |
import pandas as pd
|
| 8 |
-
from io import
|
| 9 |
from bs4 import BeautifulSoup
|
| 10 |
from duckduckgo_search import DDGS
|
| 11 |
import wikipedia
|
|
@@ -19,16 +21,17 @@ from llama_index.core.agent import ReActAgent
|
|
| 19 |
from llama_index.core.callbacks.llama_debug import LlamaDebugHandler
|
| 20 |
|
| 21 |
# -------------------------------------------------------------------
|
| 22 |
-
# 1)
|
| 23 |
# -------------------------------------------------------------------
|
| 24 |
-
# Monkey-patch para exponer .message al objeto completo, tal como espera LlamaIndex
|
| 25 |
ChatMessage.message = property(lambda self: self)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
class GeminiLLM(LLM):
|
| 27 |
model_name: str = Field(default="models/gemini-1.5-flash-latest")
|
| 28 |
temperature: float = Field(default=0.0)
|
| 29 |
|
| 30 |
-
# Atributos para el modelo y config de generación.
|
| 31 |
-
# Pydantic los ignorará si no son Fields y Config.extra = "allow" (lo cual tienes)
|
| 32 |
_model: object = None
|
| 33 |
_gen_cfg: object = None
|
| 34 |
|
|
@@ -36,78 +39,58 @@ class GeminiLLM(LLM):
|
|
| 36 |
extra = "allow"
|
| 37 |
|
| 38 |
def __init__(self, **kwargs):
|
| 39 |
-
super().__init__(**kwargs)
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
# Obtener el valor resuelto de model_name explícitamente
|
| 43 |
-
# Primero, intentar con el atributo de instancia (que Pydantic debería haber establecido)
|
| 44 |
-
actual_model_name = self.model_name
|
| 45 |
-
|
| 46 |
-
# Si sigue siendo un FieldInfo (o no es un string), obtener el valor default del campo
|
| 47 |
if not isinstance(actual_model_name, str):
|
| 48 |
-
# Acceder a la definición del campo de la clase para obtener su default
|
| 49 |
-
# self.__fields__ es un dict de los campos Pydantic de la clase
|
| 50 |
model_field_definition = self.__fields__.get("model_name")
|
| 51 |
if model_field_definition and hasattr(model_field_definition, 'default'):
|
| 52 |
actual_model_name = model_field_definition.default
|
| 53 |
-
|
| 54 |
-
# Como última salvaguarda, si todo falla, usar un string literal (no ideal)
|
| 55 |
if not isinstance(actual_model_name, str):
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
# Lo mismo para temperature, aunque es menos probable que sea un FieldInfo aquí
|
| 60 |
actual_temperature = self.temperature
|
| 61 |
if not isinstance(actual_temperature, (float, int)):
|
| 62 |
temp_field_definition = self.__fields__.get("temperature")
|
| 63 |
if temp_field_definition and hasattr(temp_field_definition, 'default'):
|
| 64 |
actual_temperature = temp_field_definition.default
|
| 65 |
if not isinstance(actual_temperature, (float, int)):
|
| 66 |
-
# print("ADVERTENCIA: temperature no se pudo resolver a un float, usando 0.0.")
|
| 67 |
actual_temperature = 0.0
|
| 68 |
-
# --- FIN DE LA CORRECCIÓN PARA FieldInfo ---
|
| 69 |
|
| 70 |
key = os.getenv("GEMINI_API_KEY")
|
| 71 |
if not key:
|
| 72 |
-
raise ValueError("GEMINI_API_KEY no configurada")
|
| 73 |
genai.configure(api_key=key)
|
| 74 |
-
|
| 75 |
self._gen_cfg = genai.types.GenerationConfig(temperature=actual_temperature)
|
| 76 |
self._model = genai.GenerativeModel(
|
| 77 |
-
model_name=actual_model_name,
|
| 78 |
generation_config=self._gen_cfg
|
| 79 |
)
|
| 80 |
-
|
| 81 |
if self.callback_manager is None:
|
| 82 |
from llama_index.core.callbacks.base import CallbackManager
|
| 83 |
self.callback_manager = CallbackManager([])
|
| 84 |
-
|
| 85 |
if not self.callback_manager.handlers:
|
| 86 |
self.callback_manager.add_handler(LlamaDebugHandler())
|
| 87 |
|
| 88 |
@property
|
| 89 |
def metadata(self):
|
| 90 |
-
# También asegurar que model_name es un string aquí
|
| 91 |
actual_model_name_meta = self.model_name
|
| 92 |
if not isinstance(actual_model_name_meta, str):
|
| 93 |
model_field_def_meta = self.__fields__.get("model_name")
|
| 94 |
if model_field_def_meta and hasattr(model_field_def_meta, 'default'):
|
| 95 |
actual_model_name_meta = model_field_def_meta.default
|
| 96 |
if not isinstance(actual_model_name_meta, str):
|
| 97 |
-
|
| 98 |
-
|
| 99 |
return LLMMetadata(
|
| 100 |
context_window=1048576,
|
| 101 |
num_output=8192,
|
| 102 |
is_chat_model=True,
|
| 103 |
is_function_calling_model=True,
|
| 104 |
-
model_name=actual_model_name_meta,
|
| 105 |
)
|
| 106 |
|
| 107 |
-
# ... (todos los demás métodos: chat, achat, stream_complete, astream_complete, stream_chat, astream_chat, complete, acomplete)
|
| 108 |
-
# DEBEN ESTAR EXACTAMENTE COMO EN TU ÚLTIMA VERSIÓN FUNCIONAL DEL CÓDIGO QUE ME PEGASTE.
|
| 109 |
-
# Los copio de tu último fragmento para asegurar consistencia:
|
| 110 |
-
|
| 111 |
def chat(self, messages: list[ChatMessage], **kwargs):
|
| 112 |
hist = []
|
| 113 |
for m in messages[:-1]:
|
|
@@ -180,20 +163,27 @@ class GeminiLLM(LLM):
|
|
| 180 |
async def acomplete(self, prompt: str, formatted=False, **kwargs):
|
| 181 |
return await asyncio.to_thread(self.complete, prompt, formatted=formatted, **kwargs)
|
| 182 |
|
| 183 |
-
# --- Fin de la clase GeminiLLM ---
|
| 184 |
# -------------------------------------------------------------------
|
| 185 |
-
#
|
| 186 |
# -------------------------------------------------------------------
|
| 187 |
HEADERS = {'User-Agent': 'Mozilla/5.0'}
|
| 188 |
|
| 189 |
-
# Web search tool
|
| 190 |
def buscar_web(query: str, max_attempts: int = 2, num_results: int = 5) -> str:
|
|
|
|
|
|
|
|
|
|
| 191 |
for i in range(max_attempts):
|
| 192 |
try:
|
| 193 |
with DDGS(headers=HEADERS, timeout=25) as ddgs:
|
| 194 |
results = list(ddgs.text(query, region='es-es', safesearch='moderate', max_results=num_results))
|
| 195 |
if results:
|
| 196 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 197 |
return "No se encontraron resultados relevantes."
|
| 198 |
except Exception as e:
|
| 199 |
if i < max_attempts - 1:
|
|
@@ -201,15 +191,19 @@ def buscar_web(query: str, max_attempts: int = 2, num_results: int = 5) -> str:
|
|
| 201 |
else:
|
| 202 |
return f"Error buscar_web tras {max_attempts} intentos: {e}"
|
| 203 |
|
| 204 |
-
# Reverse text tool
|
| 205 |
def reverse_text(text: str) -> str:
|
|
|
|
| 206 |
return text[::-1]
|
| 207 |
|
| 208 |
-
# Analyze markdown table
|
| 209 |
def analyze_table(table_md: str, question: str) -> str:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
try:
|
|
|
|
| 211 |
lines = [l for l in table_md.splitlines() if l.strip() and '---' not in l]
|
| 212 |
-
rows = [
|
| 213 |
if len(rows) < 2:
|
| 214 |
return "Tabla Markdown mal formateada o vacía."
|
| 215 |
df = pd.DataFrame(rows[1:], columns=rows[0])
|
|
@@ -219,7 +213,7 @@ def analyze_table(table_md: str, question: str) -> str:
|
|
| 219 |
for x in cols:
|
| 220 |
for y in cols:
|
| 221 |
try:
|
| 222 |
-
if df.loc[df[rows[0][0]]==x, y].iat[0] != df.loc[df[rows[0][0]]==y, x].iat[0]:
|
| 223 |
counter.update([x, y])
|
| 224 |
except:
|
| 225 |
continue
|
|
@@ -228,15 +222,18 @@ def analyze_table(table_md: str, question: str) -> str:
|
|
| 228 |
except Exception as e:
|
| 229 |
return f"Error analyze_table: {e}"
|
| 230 |
|
| 231 |
-
# Execute code tool
|
| 232 |
def execute_code(code: str) -> str:
|
|
|
|
|
|
|
|
|
|
| 233 |
try:
|
| 234 |
allowed_globals = {'__builtins__': None, 'math': math}
|
| 235 |
try:
|
| 236 |
val = eval(code, allowed_globals, {})
|
| 237 |
return str(val)
|
| 238 |
except:
|
| 239 |
-
res = subprocess.run(["python", "-S", "-c", code],
|
|
|
|
| 240 |
if res.returncode != 0:
|
| 241 |
return f"Error código: {res.stderr.strip()}"
|
| 242 |
return res.stdout.strip() or "(sin salida)"
|
|
@@ -245,10 +242,13 @@ def execute_code(code: str) -> str:
|
|
| 245 |
except Exception as e:
|
| 246 |
return f"Error crítico: {e}"
|
| 247 |
|
| 248 |
-
# Read Excel tool
|
| 249 |
def read_excel_data(file_path: str, sheet_name=0) -> str:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 250 |
try:
|
| 251 |
-
if file_path.startswith(('http://','https://')):
|
| 252 |
resp = requests.get(file_path, headers=HEADERS, timeout=30)
|
| 253 |
resp.raise_for_status()
|
| 254 |
df = pd.read_excel(BytesIO(resp.content), sheet_name=sheet_name)
|
|
@@ -262,21 +262,26 @@ def read_excel_data(file_path: str, sheet_name=0) -> str:
|
|
| 262 |
return f"Error read_excel_data: {e}"
|
| 263 |
|
| 264 |
def classify_botanical(items_list_str: str) -> str:
|
| 265 |
-
|
|
|
|
|
|
|
|
|
|
| 266 |
mapping = {
|
| 267 |
"tomato": "tomate", "pepper": "pimiento", "bell pepper": "pimiento",
|
| 268 |
"cucumber": "pepino", "eggplant": "berenjena", "zucchini": "calabacín",
|
| 269 |
"avocado": "aguacate", "squash": "calabaza", "pea": "guisante", "corn": "maíz",
|
| 270 |
-
"bean": "judía", "green beans": "judía verde", "sweet potato": "batata",
|
|
|
|
| 271 |
}
|
| 272 |
-
fruits = {"tomate","pepino","calabacín","berenjena","pimiento","aguacate","calabaza","guisante","judía verde","maíz"}
|
| 273 |
-
vegetables = {"zanahoria","patata","batata","cebolla","ajo","puerro","apio","lechuga","espinaca","brócoli","
|
| 274 |
-
|
| 275 |
items = []
|
| 276 |
for raw in items_list_str.split(','):
|
| 277 |
itm = raw.strip().lower()
|
| 278 |
itm_es = mapping.get(itm, itm)
|
| 279 |
items.append(itm_es)
|
|
|
|
| 280 |
vegs = [i for i in items if i in vegetables]
|
| 281 |
fruits_found = [i for i in items if i in fruits]
|
| 282 |
others = [i for i in items if i not in fruits and i not in vegetables]
|
|
@@ -286,116 +291,170 @@ def classify_botanical(items_list_str: str) -> str:
|
|
| 286 |
f"Otros: {', '.join(sorted(set(others)))}"
|
| 287 |
)
|
| 288 |
|
| 289 |
-
# Wikipedia table scraper
|
| 290 |
def scrape_wikipedia_table(page_title: str, section: str, table_index: int = 0) -> str:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 291 |
try:
|
| 292 |
wikipedia.set_lang("es")
|
| 293 |
page = wikipedia.page(page_title, auto_suggest=False)
|
| 294 |
soup = BeautifulSoup(page.html(), 'html.parser')
|
| 295 |
-
header = next(
|
|
|
|
|
|
|
|
|
|
| 296 |
if not header:
|
| 297 |
return f"Sección '{section}' no encontrada en '{page_title}'"
|
| 298 |
tables = []
|
| 299 |
for sib in header.find_next_siblings():
|
| 300 |
-
if sib.name in ['h2','h3']:
|
| 301 |
-
|
| 302 |
-
|
|
|
|
|
|
|
|
|
|
| 303 |
df = pd.read_html(str(tables[table_index]))[0]
|
| 304 |
return df.to_csv(index=False)
|
| 305 |
except Exception as e:
|
| 306 |
return f"Error scrape_wiki_table: {e}"
|
| 307 |
|
| 308 |
-
#
|
| 309 |
-
|
| 310 |
-
|
| 311 |
-
|
| 312 |
-
|
| 313 |
-
|
| 314 |
-
|
| 315 |
-
|
| 316 |
-
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 327 |
tool_descriptions = "\n".join([
|
| 328 |
f"{t.metadata.name}: {t.metadata.description} "
|
| 329 |
+ {
|
| 330 |
-
"classify_botanical_foods": "(
|
| 331 |
-
"read_excel_data": "(
|
| 332 |
-
"analyze_markdown_table": "(
|
| 333 |
-
"web_search": "(
|
| 334 |
-
"scrape_wiki_table": "(
|
| 335 |
-
"reverse_text": "(
|
| 336 |
-
"execute_code": "(
|
| 337 |
}.get(t.metadata.name, "")
|
| 338 |
for t in all_tools
|
| 339 |
])
|
| 340 |
|
| 341 |
-
#
|
|
|
|
|
|
|
| 342 |
system_prompt = f"""
|
| 343 |
Eres Alfred, un agente ReAct eficiente y preciso. Tu objetivo es responder correctamente usando las herramientas disponibles.
|
| 344 |
-
Sigue este flujo
|
| 345 |
-
1. LEE la pregunta y analiza palabras clave
|
| 346 |
-
- Si
|
| 347 |
-
|
| 348 |
-
- Si
|
| 349 |
-
|
| 350 |
-
- Si
|
| 351 |
-
|
| 352 |
-
- Si
|
| 353 |
-
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
3. EJECUTA la herramienta y observa el resultado.
|
| 362 |
-
4. VERIFICA que
|
| 363 |
-
5. RESPONDE de forma clara y concisa usando la salida
|
| 364 |
|
| 365 |
-
Herramientas disponibles (
|
| 366 |
{tool_descriptions}
|
| 367 |
"""
|
| 368 |
|
| 369 |
-
#
|
| 370 |
-
|
|
|
|
|
|
|
| 371 |
alfred_agent = ReActAgent.from_tools(
|
| 372 |
tools=all_tools,
|
| 373 |
llm=llm,
|
| 374 |
system_prompt=system_prompt,
|
| 375 |
verbose=True,
|
| 376 |
-
max_iterations=25,
|
| 377 |
-
callback_manager=llm.callback_manager,
|
| 378 |
-
handle_parsing_errors=True
|
| 379 |
)
|
| 380 |
|
| 381 |
def basic_agent_response(question: str) -> str:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 382 |
try:
|
| 383 |
-
|
| 384 |
-
|
| 385 |
-
|
| 386 |
-
# Supongamos que SAIA añade algo como “see attached Excel file” sin ruta en la pregunta.
|
| 387 |
-
# Podemos devolver un mensaje especial que indique al usuario que suba la ruta.
|
| 388 |
-
# Pero en la práctica, SAIA pasa el path en un campo aparte; aquí solo forzamos a usar read_excel_data:
|
| 389 |
-
# (En muchos casos, SAIA evalúa que invoques la herramienta correctamente)
|
| 390 |
-
# Podemos invocar read_excel_data sin parámetros y devolver un placeholder:
|
| 391 |
return read_excel_data("data/attached.xlsx")
|
| 392 |
-
|
| 393 |
resp = alfred_agent.query(question)
|
| 394 |
if hasattr(resp, 'response') and resp.response is not None:
|
| 395 |
return str(resp.response)
|
| 396 |
elif resp is not None:
|
| 397 |
return str(resp)
|
| 398 |
-
|
|
|
|
| 399 |
except Exception as e:
|
| 400 |
return f"Error crítico del agente: {e}"
|
| 401 |
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# my_tools.py
|
| 2 |
+
|
| 3 |
import os
|
| 4 |
import math
|
| 5 |
import time
|
|
|
|
| 7 |
import subprocess
|
| 8 |
import requests
|
| 9 |
import pandas as pd
|
| 10 |
+
from io import BytesIO
|
| 11 |
from bs4 import BeautifulSoup
|
| 12 |
from duckduckgo_search import DDGS
|
| 13 |
import wikipedia
|
|
|
|
| 21 |
from llama_index.core.callbacks.llama_debug import LlamaDebugHandler
|
| 22 |
|
| 23 |
# -------------------------------------------------------------------
|
| 24 |
+
# 1) MONKEY-PATCH PARA ChatMessage (por requerimiento de LlamaIndex)
|
| 25 |
# -------------------------------------------------------------------
|
|
|
|
| 26 |
ChatMessage.message = property(lambda self: self)
|
| 27 |
+
|
| 28 |
+
# -------------------------------------------------------------------
|
| 29 |
+
# 2) Clase GeminiLLM personalizada
|
| 30 |
+
# -------------------------------------------------------------------
|
| 31 |
class GeminiLLM(LLM):
|
| 32 |
model_name: str = Field(default="models/gemini-1.5-flash-latest")
|
| 33 |
temperature: float = Field(default=0.0)
|
| 34 |
|
|
|
|
|
|
|
| 35 |
_model: object = None
|
| 36 |
_gen_cfg: object = None
|
| 37 |
|
|
|
|
| 39 |
extra = "allow"
|
| 40 |
|
| 41 |
def __init__(self, **kwargs):
|
| 42 |
+
super().__init__(**kwargs)
|
| 43 |
+
# Resolver FieldInfo si es necesario
|
| 44 |
+
actual_model_name = self.model_name
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
if not isinstance(actual_model_name, str):
|
|
|
|
|
|
|
| 46 |
model_field_definition = self.__fields__.get("model_name")
|
| 47 |
if model_field_definition and hasattr(model_field_definition, 'default'):
|
| 48 |
actual_model_name = model_field_definition.default
|
|
|
|
|
|
|
| 49 |
if not isinstance(actual_model_name, str):
|
| 50 |
+
actual_model_name = "models/gemini-1.5-flash-latest"
|
| 51 |
+
|
|
|
|
|
|
|
| 52 |
actual_temperature = self.temperature
|
| 53 |
if not isinstance(actual_temperature, (float, int)):
|
| 54 |
temp_field_definition = self.__fields__.get("temperature")
|
| 55 |
if temp_field_definition and hasattr(temp_field_definition, 'default'):
|
| 56 |
actual_temperature = temp_field_definition.default
|
| 57 |
if not isinstance(actual_temperature, (float, int)):
|
|
|
|
| 58 |
actual_temperature = 0.0
|
|
|
|
| 59 |
|
| 60 |
key = os.getenv("GEMINI_API_KEY")
|
| 61 |
if not key:
|
| 62 |
+
raise ValueError("GEMINI_API_KEY no configurada en variables de entorno")
|
| 63 |
genai.configure(api_key=key)
|
| 64 |
+
|
| 65 |
self._gen_cfg = genai.types.GenerationConfig(temperature=actual_temperature)
|
| 66 |
self._model = genai.GenerativeModel(
|
| 67 |
+
model_name=actual_model_name,
|
| 68 |
generation_config=self._gen_cfg
|
| 69 |
)
|
| 70 |
+
|
| 71 |
if self.callback_manager is None:
|
| 72 |
from llama_index.core.callbacks.base import CallbackManager
|
| 73 |
self.callback_manager = CallbackManager([])
|
|
|
|
| 74 |
if not self.callback_manager.handlers:
|
| 75 |
self.callback_manager.add_handler(LlamaDebugHandler())
|
| 76 |
|
| 77 |
@property
|
| 78 |
def metadata(self):
|
|
|
|
| 79 |
actual_model_name_meta = self.model_name
|
| 80 |
if not isinstance(actual_model_name_meta, str):
|
| 81 |
model_field_def_meta = self.__fields__.get("model_name")
|
| 82 |
if model_field_def_meta and hasattr(model_field_def_meta, 'default'):
|
| 83 |
actual_model_name_meta = model_field_def_meta.default
|
| 84 |
if not isinstance(actual_model_name_meta, str):
|
| 85 |
+
actual_model_name_meta = "models/gemini-1.5-flash-latest"
|
|
|
|
| 86 |
return LLMMetadata(
|
| 87 |
context_window=1048576,
|
| 88 |
num_output=8192,
|
| 89 |
is_chat_model=True,
|
| 90 |
is_function_calling_model=True,
|
| 91 |
+
model_name=actual_model_name_meta,
|
| 92 |
)
|
| 93 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
def chat(self, messages: list[ChatMessage], **kwargs):
|
| 95 |
hist = []
|
| 96 |
for m in messages[:-1]:
|
|
|
|
| 163 |
async def acomplete(self, prompt: str, formatted=False, **kwargs):
|
| 164 |
return await asyncio.to_thread(self.complete, prompt, formatted=formatted, **kwargs)
|
| 165 |
|
|
|
|
| 166 |
# -------------------------------------------------------------------
|
| 167 |
+
# 3) HERRAMIENTAS PERSONALIZADAS
|
| 168 |
# -------------------------------------------------------------------
|
| 169 |
HEADERS = {'User-Agent': 'Mozilla/5.0'}
|
| 170 |
|
|
|
|
| 171 |
def buscar_web(query: str, max_attempts: int = 2, num_results: int = 5) -> str:
|
| 172 |
+
"""
|
| 173 |
+
Usa DuckDuckGo (vía duckduckgo_search) para devolver hasta 'num_results' resultados.
|
| 174 |
+
"""
|
| 175 |
for i in range(max_attempts):
|
| 176 |
try:
|
| 177 |
with DDGS(headers=HEADERS, timeout=25) as ddgs:
|
| 178 |
results = list(ddgs.text(query, region='es-es', safesearch='moderate', max_results=num_results))
|
| 179 |
if results:
|
| 180 |
+
salida = []
|
| 181 |
+
for idx, r in enumerate(results):
|
| 182 |
+
titulo = r.get('title', 'Sin título')
|
| 183 |
+
enlace = r.get('href', 'N/A')
|
| 184 |
+
cuerpo = r.get('body', '')
|
| 185 |
+
salida.append(f"Fuente {idx+1}: Título: {titulo}\nEnlace: {enlace}\nCuerpo: {cuerpo}")
|
| 186 |
+
return "\n\n".join(salida)
|
| 187 |
return "No se encontraron resultados relevantes."
|
| 188 |
except Exception as e:
|
| 189 |
if i < max_attempts - 1:
|
|
|
|
| 191 |
else:
|
| 192 |
return f"Error buscar_web tras {max_attempts} intentos: {e}"
|
| 193 |
|
|
|
|
| 194 |
def reverse_text(text: str) -> str:
|
| 195 |
+
"""Invierte el orden de los caracteres en 'text'."""
|
| 196 |
return text[::-1]
|
| 197 |
|
|
|
|
| 198 |
def analyze_table(table_md: str, question: str) -> str:
|
| 199 |
+
"""
|
| 200 |
+
Recibe una tabla en Markdown (con pipes y separadores) y, si la pregunta menciona 'conmut',
|
| 201 |
+
verifica la conmutatividad de la matriz; en otro caso, devuelve el CSV equivalente.
|
| 202 |
+
"""
|
| 203 |
try:
|
| 204 |
+
# Quitar líneas de separación y vacías
|
| 205 |
lines = [l for l in table_md.splitlines() if l.strip() and '---' not in l]
|
| 206 |
+
rows = [[c.strip() for c in l.strip().strip('|').split('|')] for l in lines]
|
| 207 |
if len(rows) < 2:
|
| 208 |
return "Tabla Markdown mal formateada o vacía."
|
| 209 |
df = pd.DataFrame(rows[1:], columns=rows[0])
|
|
|
|
| 213 |
for x in cols:
|
| 214 |
for y in cols:
|
| 215 |
try:
|
| 216 |
+
if df.loc[df[rows[0][0]] == x, y].iat[0] != df.loc[df[rows[0][0]] == y, x].iat[0]:
|
| 217 |
counter.update([x, y])
|
| 218 |
except:
|
| 219 |
continue
|
|
|
|
| 222 |
except Exception as e:
|
| 223 |
return f"Error analyze_table: {e}"
|
| 224 |
|
|
|
|
| 225 |
def execute_code(code: str) -> str:
|
| 226 |
+
"""
|
| 227 |
+
Primero intenta evaluar con eval() en un entorno protegido; si falla, invoca un subproceso 'python -c'.
|
| 228 |
+
"""
|
| 229 |
try:
|
| 230 |
allowed_globals = {'__builtins__': None, 'math': math}
|
| 231 |
try:
|
| 232 |
val = eval(code, allowed_globals, {})
|
| 233 |
return str(val)
|
| 234 |
except:
|
| 235 |
+
res = subprocess.run(["python", "-S", "-c", code],
|
| 236 |
+
capture_output=True, text=True, timeout=10)
|
| 237 |
if res.returncode != 0:
|
| 238 |
return f"Error código: {res.stderr.strip()}"
|
| 239 |
return res.stdout.strip() or "(sin salida)"
|
|
|
|
| 242 |
except Exception as e:
|
| 243 |
return f"Error crítico: {e}"
|
| 244 |
|
|
|
|
| 245 |
def read_excel_data(file_path: str, sheet_name=0) -> str:
|
| 246 |
+
"""
|
| 247 |
+
Si file_path empieza con 'http', descarga el contenido y lee con pandas.
|
| 248 |
+
Si es una ruta local, lee directamente. Devuelve el CSV.
|
| 249 |
+
"""
|
| 250 |
try:
|
| 251 |
+
if file_path.startswith(('http://', 'https://')):
|
| 252 |
resp = requests.get(file_path, headers=HEADERS, timeout=30)
|
| 253 |
resp.raise_for_status()
|
| 254 |
df = pd.read_excel(BytesIO(resp.content), sheet_name=sheet_name)
|
|
|
|
| 262 |
return f"Error read_excel_data: {e}"
|
| 263 |
|
| 264 |
def classify_botanical(items_list_str: str) -> str:
|
| 265 |
+
"""
|
| 266 |
+
Clasifica botánicamente una lista de alimentos (en inglés o español) en Verduras, Frutas u Otros.
|
| 267 |
+
"""
|
| 268 |
+
# Mapeo inglés → español
|
| 269 |
mapping = {
|
| 270 |
"tomato": "tomate", "pepper": "pimiento", "bell pepper": "pimiento",
|
| 271 |
"cucumber": "pepino", "eggplant": "berenjena", "zucchini": "calabacín",
|
| 272 |
"avocado": "aguacate", "squash": "calabaza", "pea": "guisante", "corn": "maíz",
|
| 273 |
+
"bean": "judía", "beans": "judía", "green beans": "judía verde", "sweet potato": "batata",
|
| 274 |
+
"whole bean coffee": "café", "rice": "arroz", "oregano": "orégano"
|
| 275 |
}
|
| 276 |
+
fruits = {"tomate", "pepino", "calabacín", "berenjena", "pimiento", "aguacate", "calabaza", "guisante", "judía verde", "maíz"}
|
| 277 |
+
vegetables = {"zanahoria", "patata", "batata", "cebolla", "ajo", "puerro", "apio", "lechuga", "espinaca", "brócoli", "pepino", "pepino"}
|
| 278 |
+
|
| 279 |
items = []
|
| 280 |
for raw in items_list_str.split(','):
|
| 281 |
itm = raw.strip().lower()
|
| 282 |
itm_es = mapping.get(itm, itm)
|
| 283 |
items.append(itm_es)
|
| 284 |
+
|
| 285 |
vegs = [i for i in items if i in vegetables]
|
| 286 |
fruits_found = [i for i in items if i in fruits]
|
| 287 |
others = [i for i in items if i not in fruits and i not in vegetables]
|
|
|
|
| 291 |
f"Otros: {', '.join(sorted(set(others)))}"
|
| 292 |
)
|
| 293 |
|
|
|
|
| 294 |
def scrape_wikipedia_table(page_title: str, section: str, table_index: int = 0) -> str:
|
| 295 |
+
"""
|
| 296 |
+
Busca una sección en una página de Wikipedia y extrae la tabla indicada (por índice).
|
| 297 |
+
Devuelve el CSV.
|
| 298 |
+
"""
|
| 299 |
try:
|
| 300 |
wikipedia.set_lang("es")
|
| 301 |
page = wikipedia.page(page_title, auto_suggest=False)
|
| 302 |
soup = BeautifulSoup(page.html(), 'html.parser')
|
| 303 |
+
header = next(
|
| 304 |
+
(h for h in soup.find_all(['h2', 'h3']) if section.lower() in h.get_text(strip=True).lower()),
|
| 305 |
+
None
|
| 306 |
+
)
|
| 307 |
if not header:
|
| 308 |
return f"Sección '{section}' no encontrada en '{page_title}'"
|
| 309 |
tables = []
|
| 310 |
for sib in header.find_next_siblings():
|
| 311 |
+
if sib.name in ['h2', 'h3']:
|
| 312 |
+
break
|
| 313 |
+
if sib.name == 'table' and 'wikitable' in sib.get('class', []):
|
| 314 |
+
tables.append(sib)
|
| 315 |
+
if table_index >= len(tables):
|
| 316 |
+
return f"Tabla índice {table_index} fuera de rango (solo {len(tables)} tablas)."
|
| 317 |
df = pd.read_html(str(tables[table_index]))[0]
|
| 318 |
return df.to_csv(index=False)
|
| 319 |
except Exception as e:
|
| 320 |
return f"Error scrape_wiki_table: {e}"
|
| 321 |
|
| 322 |
+
# -------------------------------------------------------------------
|
| 323 |
+
# 4) ENVUELTORES DE HERRAMIENTAS (FunctionTool)
|
| 324 |
+
# -------------------------------------------------------------------
|
| 325 |
+
search_tool = FunctionTool.from_defaults(
|
| 326 |
+
fn=buscar_web,
|
| 327 |
+
name="web_search",
|
| 328 |
+
description="Búsqueda DuckDuckGo (máximo 5 resultados)."
|
| 329 |
+
)
|
| 330 |
+
reverse_tool = FunctionTool.from_defaults(
|
| 331 |
+
fn=reverse_text,
|
| 332 |
+
name="reverse_text",
|
| 333 |
+
description="Invierte el texto recibido."
|
| 334 |
+
)
|
| 335 |
+
table_tool = FunctionTool.from_defaults(
|
| 336 |
+
fn=analyze_table,
|
| 337 |
+
name="analyze_markdown_table",
|
| 338 |
+
description="Procesa tabla Markdown y verifica conmutatividad si se menciona 'conmut'."
|
| 339 |
+
)
|
| 340 |
+
code_tool = FunctionTool.from_defaults(
|
| 341 |
+
fn=execute_code,
|
| 342 |
+
name="execute_code",
|
| 343 |
+
description="Ejecuta código Python de forma segura."
|
| 344 |
+
)
|
| 345 |
+
excel_tool = FunctionTool.from_defaults(
|
| 346 |
+
fn=read_excel_data,
|
| 347 |
+
name="read_excel_data",
|
| 348 |
+
description="Lee un archivo Excel (local o URL) y devuelve CSV."
|
| 349 |
+
)
|
| 350 |
+
botanical_tool = FunctionTool.from_defaults(
|
| 351 |
+
fn=classify_botanical,
|
| 352 |
+
name="classify_botanical_foods",
|
| 353 |
+
description="Clasifica botánicamente una lista de alimentos."
|
| 354 |
+
)
|
| 355 |
+
scrape_tool = FunctionTool.from_defaults(
|
| 356 |
+
fn=scrape_wikipedia_table,
|
| 357 |
+
name="scrape_wiki_table",
|
| 358 |
+
description="Extrae tabla de sección específica de Wikipedia."
|
| 359 |
+
)
|
| 360 |
+
fallback_tool = FunctionTool.from_defaults(
|
| 361 |
+
fn=lambda q: "Procedo con conocimiento interno.",
|
| 362 |
+
name="no_tool_solution",
|
| 363 |
+
description="Respuesta genérica por conocimiento interno si todo lo demás falla."
|
| 364 |
+
)
|
| 365 |
+
|
| 366 |
+
all_tools = [
|
| 367 |
+
search_tool,
|
| 368 |
+
scrape_tool,
|
| 369 |
+
table_tool,
|
| 370 |
+
code_tool,
|
| 371 |
+
excel_tool,
|
| 372 |
+
botanical_tool,
|
| 373 |
+
reverse_tool,
|
| 374 |
+
fallback_tool
|
| 375 |
+
]
|
| 376 |
+
|
| 377 |
+
# -------------------------------------------------------------------
|
| 378 |
+
# 5) DESCRIPCIONES DE HERRAMIENTAS (con ejemplos)
|
| 379 |
+
# -------------------------------------------------------------------
|
| 380 |
tool_descriptions = "\n".join([
|
| 381 |
f"{t.metadata.name}: {t.metadata.description} "
|
| 382 |
+ {
|
| 383 |
+
"classify_botanical_foods": "(Ej: classify_botanical_foods('zanahoria, pepino, tomate'))",
|
| 384 |
+
"read_excel_data": "(Ej: read_excel_data('ventas.xlsx', sheet_name=0))",
|
| 385 |
+
"analyze_markdown_table": "(Ej: analyze_markdown_table('| A | B |\\n|---|---|\\n|1|2|', '¿Es conmut?'))",
|
| 386 |
+
"web_search": "(Ej: web_search('¿Quién ganó la Champions 2025?'))",
|
| 387 |
+
"scrape_wiki_table": "(Ej: scrape_wiki_table('Lionel Messi', 'Carrera', 0))",
|
| 388 |
+
"reverse_text": "(Ej: reverse_text('Hola'))",
|
| 389 |
+
"execute_code": "(Ej: execute_code('5*7'))",
|
| 390 |
}.get(t.metadata.name, "")
|
| 391 |
for t in all_tools
|
| 392 |
])
|
| 393 |
|
| 394 |
+
# -------------------------------------------------------------------
|
| 395 |
+
# 6) PROMPT DE SISTEMA MEJORADO
|
| 396 |
+
# -------------------------------------------------------------------
|
| 397 |
system_prompt = f"""
|
| 398 |
Eres Alfred, un agente ReAct eficiente y preciso. Tu objetivo es responder correctamente usando las herramientas disponibles.
|
| 399 |
+
Sigue este flujo en cada pregunta:
|
| 400 |
+
1. LEE la pregunta y analiza palabras clave:
|
| 401 |
+
- Si menciona “lista” de “alimentos” o “categorizar” botánicamente, llama:
|
| 402 |
+
classify_botanical_foods(<lista_coma_sep>).
|
| 403 |
+
- Si menciona “archivo Excel” o “Excel adjunto”, llama:
|
| 404 |
+
read_excel_data(<ruta_o_URL>).
|
| 405 |
+
- Si ves una “tabla Markdown”, llama:
|
| 406 |
+
analyze_markdown_table(<tabla_md>, <pregunta>).
|
| 407 |
+
- Si necesitas información general de la web, llama:
|
| 408 |
+
web_search(<consulta>).
|
| 409 |
+
- Si necesitas raspar tablas de Wikipedia, llama:
|
| 410 |
+
scrape_wiki_table(<título>, <sección>, <índice_tabla>).
|
| 411 |
+
- Si hay que invertir texto, llama:
|
| 412 |
+
reverse_text(<texto>).
|
| 413 |
+
- Si hay que ejecutar código Python, llama:
|
| 414 |
+
execute_code(<código>).
|
| 415 |
+
2. GENERA el “TOOL CALL” con la entrada correcta.
|
| 416 |
3. EJECUTA la herramienta y observa el resultado.
|
| 417 |
+
4. VERIFICA que el resultado responda bien la pregunta. Si no, intenta otro paso.
|
| 418 |
+
5. RESPONDE de forma clara y concisa usando la salida obtenida.
|
| 419 |
|
| 420 |
+
Herramientas disponibles (USAR EXÁCTAMENTE estos nombres):
|
| 421 |
{tool_descriptions}
|
| 422 |
"""
|
| 423 |
|
| 424 |
+
# -------------------------------------------------------------------
|
| 425 |
+
# 7) INICIALIZAR EL AGENTE ReActAgent
|
| 426 |
+
# -------------------------------------------------------------------
|
| 427 |
+
llm = GeminiLLM()
|
| 428 |
alfred_agent = ReActAgent.from_tools(
|
| 429 |
tools=all_tools,
|
| 430 |
llm=llm,
|
| 431 |
system_prompt=system_prompt,
|
| 432 |
verbose=True,
|
| 433 |
+
max_iterations=25, # Más iteraciones para razonamiento multi-paso
|
| 434 |
+
callback_manager=llm.callback_manager,
|
| 435 |
+
handle_parsing_errors=True # Para que reintente si la llamada a herramienta sale malformada
|
| 436 |
)
|
| 437 |
|
| 438 |
def basic_agent_response(question: str) -> str:
|
| 439 |
+
"""
|
| 440 |
+
Si detecta “Excel adjunto”, asume que SAIA inyecta el path y fuerza read_excel_data.
|
| 441 |
+
De lo contrario, usa ReActAgent.query().
|
| 442 |
+
"""
|
| 443 |
try:
|
| 444 |
+
# Forzar uso de read_excel_data si aparece Excel en la pregunta
|
| 445 |
+
if "attached excel" in question.lower() or "archivo excel" in question.lower():
|
| 446 |
+
# En el entorno SAIA normalmente inyectan la ruta real; aquí usamos un placeholder.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 447 |
return read_excel_data("data/attached.xlsx")
|
|
|
|
| 448 |
resp = alfred_agent.query(question)
|
| 449 |
if hasattr(resp, 'response') and resp.response is not None:
|
| 450 |
return str(resp.response)
|
| 451 |
elif resp is not None:
|
| 452 |
return str(resp)
|
| 453 |
+
else:
|
| 454 |
+
return "No se generó una respuesta válida."
|
| 455 |
except Exception as e:
|
| 456 |
return f"Error crítico del agente: {e}"
|
| 457 |
|
| 458 |
+
# --- FIN DE my_tools.py ---
|
| 459 |
+
|
| 460 |
+
|