Update app.py
Browse files
app.py
CHANGED
|
@@ -2,139 +2,163 @@ import os
|
|
| 2 |
import re
|
| 3 |
import pdfplumber
|
| 4 |
import gradio as gr
|
| 5 |
-
from openai import OpenAI
|
| 6 |
-
from huggingface_hub import hf_hub_download, list_repo_files
|
| 7 |
from dotenv import load_dotenv
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
|
| 9 |
# ------------------------------------------------------------
|
| 10 |
-
#
|
| 11 |
# ------------------------------------------------------------
|
| 12 |
load_dotenv()
|
| 13 |
-
client = OpenAI(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
# ------------------------------------------------------------
|
| 16 |
-
#
|
| 17 |
# ------------------------------------------------------------
|
| 18 |
system_prompt = """
|
| 19 |
-
Eres un Asistente de
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
evaluación y documentación de auditorías, así como en la preparación para el
|
| 25 |
-
examen CIA (Certified Internal Auditor). Tus respuestas deben reflejar:
|
| 26 |
-
- Objetividad, integridad y confidencialidad.
|
| 27 |
-
- Los valores de Gentera: Responsabilidad, Empatía, Innovación y Transparencia.
|
| 28 |
-
- Lenguaje claro, profesional y humano.
|
| 29 |
-
|
| 30 |
-
Si la pregunta se relaciona con auditoría, control interno, riesgos o ética profesional,
|
| 31 |
-
responde con rigor técnico y ejemplos prácticos. Si se pide un resumen de un PDF,
|
| 32 |
-
integra el contenido del documento correspondiente.
|
| 33 |
"""
|
| 34 |
|
| 35 |
# ------------------------------------------------------------
|
| 36 |
-
# CARGA DE PDFs
|
| 37 |
# ------------------------------------------------------------
|
| 38 |
REPO_ID = "vecervantes89/auditoria_interna_pdfs"
|
| 39 |
REPO_TYPE = "dataset"
|
| 40 |
|
| 41 |
def extract_pdf_text(local_path: str) -> str:
|
| 42 |
-
|
| 43 |
with pdfplumber.open(local_path) as pdf:
|
| 44 |
-
for
|
| 45 |
-
|
| 46 |
-
return "\n".join(
|
| 47 |
|
| 48 |
def load_hf_pdfs_text(repo_id: str, repo_type: str = "dataset"):
|
| 49 |
try:
|
| 50 |
files = [f for f in list_repo_files(repo_id=repo_id, repo_type=repo_type) if f.lower().endswith(".pdf")]
|
| 51 |
except Exception as e:
|
| 52 |
-
print(f"[ERROR] No se pudo listar
|
| 53 |
-
return {"files": [], "
|
| 54 |
|
| 55 |
entries = []
|
| 56 |
for f in files:
|
| 57 |
try:
|
| 58 |
-
|
| 59 |
-
text = extract_pdf_text(
|
| 60 |
entries.append({"name": f, "text": text})
|
| 61 |
print(f"[OK] Cargado {f}")
|
| 62 |
except Exception as e:
|
| 63 |
print(f"[ERROR] Falló la carga de {f}: {e}")
|
| 64 |
|
| 65 |
-
all_text = "\n\n".join(e["text"] for e in entries)
|
| 66 |
by_name = {e["name"]: e["text"] for e in entries}
|
| 67 |
-
|
| 68 |
-
|
|
|
|
| 69 |
|
| 70 |
HF_DOCS = load_hf_pdfs_text(REPO_ID, REPO_TYPE)
|
| 71 |
|
| 72 |
# ------------------------------------------------------------
|
| 73 |
-
#
|
| 74 |
# ------------------------------------------------------------
|
| 75 |
def buscar_mejor_fragmento(pregunta: str, docs: dict, max_chars: int = 3000):
|
| 76 |
-
q = pregunta.lower()
|
| 77 |
-
|
| 78 |
-
# 1) Coincidencia por nombre de archivo mencionado en la pregunta
|
| 79 |
for name, text in docs.get("by_name", {}).items():
|
| 80 |
if name.lower() in q:
|
| 81 |
return name, (text or "")[:max_chars]
|
| 82 |
|
| 83 |
-
# 2) Coincidencia
|
| 84 |
tokens = [t for t in re.findall(r"[a-záéíóúüñ0-9]+", q) if len(t) > 2]
|
| 85 |
best_name, best_score, best_text = "", 0, ""
|
| 86 |
for e in docs.get("files", []):
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
if
|
| 90 |
-
best_score, best_name, best_text =
|
| 91 |
return (best_name, (best_text or "")[:max_chars]) if best_score > 0 else ("", "")
|
| 92 |
|
|
|
|
|
|
|
|
|
|
| 93 |
def responder(user_text: str, history: list | None):
|
| 94 |
-
"""
|
| 95 |
-
Con Chatbot(type="messages"), history es una lista de dicts:
|
| 96 |
-
[{"role":"user","content":"..."}, {"role":"assistant","content":"..."}]
|
| 97 |
-
"""
|
| 98 |
try:
|
| 99 |
-
history = history or []
|
| 100 |
-
# Añadimos el mensaje del usuario al historial
|
| 101 |
history.append({"role": "user", "content": user_text})
|
| 102 |
|
| 103 |
-
#
|
| 104 |
nombre_pdf, fragmento = buscar_mejor_fragmento(user_text, HF_DOCS)
|
| 105 |
if fragmento:
|
| 106 |
contenido_usuario = (
|
| 107 |
f"El siguiente texto proviene del documento '{nombre_pdf}'. "
|
| 108 |
-
"Úsalo
|
| 109 |
-
f"{fragmento}\n\
|
|
|
|
| 110 |
)
|
| 111 |
else:
|
| 112 |
contenido_usuario = user_text
|
| 113 |
|
|
|
|
| 114 |
mensajes = [{"role": "system", "content": system_prompt}] + history[:-1] + [
|
| 115 |
{"role": "user", "content": contenido_usuario}
|
| 116 |
]
|
| 117 |
|
|
|
|
| 118 |
resp = client.chat.completions.create(
|
| 119 |
model="gpt-4o-mini",
|
| 120 |
messages=mensajes,
|
| 121 |
temperature=0.3,
|
| 122 |
)
|
| 123 |
-
|
| 124 |
-
history.append({"role": "assistant", "content":
|
| 125 |
-
|
| 126 |
-
# Limpiar textbox (""), devolver historial en formato messages
|
| 127 |
return "", history
|
| 128 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 129 |
except Exception as e:
|
| 130 |
-
history.append({"role": "assistant", "content": f"⚠️ Error: {e}"})
|
| 131 |
return "", history
|
| 132 |
|
| 133 |
def limpiar_chat():
|
| 134 |
-
return [] #
|
| 135 |
|
| 136 |
# ------------------------------------------------------------
|
| 137 |
-
#
|
| 138 |
# ------------------------------------------------------------
|
| 139 |
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue")) as demo:
|
| 140 |
gr.HTML("""
|
|
@@ -144,9 +168,11 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue")) as demo:
|
|
| 144 |
<p style="font-size:15px;">Basado en GPT-4o y los valores del IIA y Gentera</p>
|
| 145 |
</div>
|
| 146 |
""")
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
|
|
|
|
|
|
| 150 |
msg = gr.Textbox(placeholder="Escribe tu consulta aquí...", label="Tu mensaje")
|
| 151 |
clear = gr.Button("🧹 Limpiar chat")
|
| 152 |
|
|
|
|
| 2 |
import re
|
| 3 |
import pdfplumber
|
| 4 |
import gradio as gr
|
|
|
|
|
|
|
| 5 |
from dotenv import load_dotenv
|
| 6 |
+
from huggingface_hub import hf_hub_download, list_repo_files
|
| 7 |
+
from openai import OpenAI
|
| 8 |
+
|
| 9 |
+
# --- Excepciones (compatibles con distintas versiones del SDK) ---
|
| 10 |
+
try:
|
| 11 |
+
from openai import (
|
| 12 |
+
APIConnectionError as _APIConnectionError,
|
| 13 |
+
APIStatusError as _APIStatusError,
|
| 14 |
+
RateLimitError as _RateLimitError,
|
| 15 |
+
AuthenticationError as _AuthenticationError,
|
| 16 |
+
APITimeoutError as _APITimeoutError,
|
| 17 |
+
)
|
| 18 |
+
except Exception:
|
| 19 |
+
_APIConnectionError = _APIStatusError = _RateLimitError = _AuthenticationError = _APITimeoutError = Exception
|
| 20 |
|
| 21 |
# ------------------------------------------------------------
|
| 22 |
+
# CONFIG: OpenAI
|
| 23 |
# ------------------------------------------------------------
|
| 24 |
load_dotenv()
|
| 25 |
+
client = OpenAI(
|
| 26 |
+
api_key=os.getenv("OPENAI_API_KEY"),
|
| 27 |
+
timeout=30, # evita cuelgues largos
|
| 28 |
+
max_retries=1, # sin reintentos largos
|
| 29 |
+
)
|
| 30 |
|
| 31 |
# ------------------------------------------------------------
|
| 32 |
+
# SYSTEM PROMPT
|
| 33 |
# ------------------------------------------------------------
|
| 34 |
system_prompt = """
|
| 35 |
+
Eres un Asistente de IA especializado en Auditoría Interna,
|
| 36 |
+
conforme a las Normas del IIA. Apoyas en análisis, planeación,
|
| 37 |
+
ejecución y documentación de auditorías y en la preparación para el CIA.
|
| 38 |
+
Responde con rigor técnico, ejemplos claros y lenguaje profesional.
|
| 39 |
+
Si la consulta menciona un PDF, integra fragmentos pertinentes del documento.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
"""
|
| 41 |
|
| 42 |
# ------------------------------------------------------------
|
| 43 |
+
# CARGA DE PDFs (dataset en Hugging Face)
|
| 44 |
# ------------------------------------------------------------
|
| 45 |
REPO_ID = "vecervantes89/auditoria_interna_pdfs"
|
| 46 |
REPO_TYPE = "dataset"
|
| 47 |
|
| 48 |
def extract_pdf_text(local_path: str) -> str:
|
| 49 |
+
parts = []
|
| 50 |
with pdfplumber.open(local_path) as pdf:
|
| 51 |
+
for p in pdf.pages:
|
| 52 |
+
parts.append(p.extract_text() or "")
|
| 53 |
+
return "\n".join(parts)
|
| 54 |
|
| 55 |
def load_hf_pdfs_text(repo_id: str, repo_type: str = "dataset"):
|
| 56 |
try:
|
| 57 |
files = [f for f in list_repo_files(repo_id=repo_id, repo_type=repo_type) if f.lower().endswith(".pdf")]
|
| 58 |
except Exception as e:
|
| 59 |
+
print(f"[ERROR] No se pudo listar '{repo_id}': {e}")
|
| 60 |
+
return {"files": [], "by_name": {}, "all_text": ""}
|
| 61 |
|
| 62 |
entries = []
|
| 63 |
for f in files:
|
| 64 |
try:
|
| 65 |
+
path = hf_hub_download(repo_id=repo_id, filename=f, repo_type=repo_type)
|
| 66 |
+
text = extract_pdf_text(path)
|
| 67 |
entries.append({"name": f, "text": text})
|
| 68 |
print(f"[OK] Cargado {f}")
|
| 69 |
except Exception as e:
|
| 70 |
print(f"[ERROR] Falló la carga de {f}: {e}")
|
| 71 |
|
|
|
|
| 72 |
by_name = {e["name"]: e["text"] for e in entries}
|
| 73 |
+
all_text = "\n\n".join(e["text"] for e in entries)
|
| 74 |
+
print(f"[INFO] Se cargaron {len(entries)} PDFs desde {repo_id}.")
|
| 75 |
+
return {"files": entries, "by_name": by_name, "all_text": all_text}
|
| 76 |
|
| 77 |
HF_DOCS = load_hf_pdfs_text(REPO_ID, REPO_TYPE)
|
| 78 |
|
| 79 |
# ------------------------------------------------------------
|
| 80 |
+
# BÚSQUEDA SIMPLE DE CONTEXTO
|
| 81 |
# ------------------------------------------------------------
|
| 82 |
def buscar_mejor_fragmento(pregunta: str, docs: dict, max_chars: int = 3000):
|
| 83 |
+
q = (pregunta or "").lower()
|
| 84 |
+
# 1) Si menciona explícitamente un nombre de archivo
|
|
|
|
| 85 |
for name, text in docs.get("by_name", {}).items():
|
| 86 |
if name.lower() in q:
|
| 87 |
return name, (text or "")[:max_chars]
|
| 88 |
|
| 89 |
+
# 2) Coincidencia por términos
|
| 90 |
tokens = [t for t in re.findall(r"[a-záéíóúüñ0-9]+", q) if len(t) > 2]
|
| 91 |
best_name, best_score, best_text = "", 0, ""
|
| 92 |
for e in docs.get("files", []):
|
| 93 |
+
t = (e.get("text") or "").lower()
|
| 94 |
+
s = sum(t.count(tok) for tok in tokens)
|
| 95 |
+
if s > best_score:
|
| 96 |
+
best_score, best_name, best_text = s, e.get("name", ""), e.get("text", "")
|
| 97 |
return (best_name, (best_text or "")[:max_chars]) if best_score > 0 else ("", "")
|
| 98 |
|
| 99 |
+
# ------------------------------------------------------------
|
| 100 |
+
# HANDLER DEL CHAT (type="messages")
|
| 101 |
+
# ------------------------------------------------------------
|
| 102 |
def responder(user_text: str, history: list | None):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
try:
|
| 104 |
+
history = history or []
|
|
|
|
| 105 |
history.append({"role": "user", "content": user_text})
|
| 106 |
|
| 107 |
+
# Contexto desde PDFs
|
| 108 |
nombre_pdf, fragmento = buscar_mejor_fragmento(user_text, HF_DOCS)
|
| 109 |
if fragmento:
|
| 110 |
contenido_usuario = (
|
| 111 |
f"El siguiente texto proviene del documento '{nombre_pdf}'. "
|
| 112 |
+
"Úsalo como contexto y responde de forma clara, breve y profesional:\n\n"
|
| 113 |
+
f"{fragmento}\n\n"
|
| 114 |
+
f"Pregunta del usuario:\n{user_text}"
|
| 115 |
)
|
| 116 |
else:
|
| 117 |
contenido_usuario = user_text
|
| 118 |
|
| 119 |
+
# Construimos los mensajes: system + historial (sin el último) + user contextualizado
|
| 120 |
mensajes = [{"role": "system", "content": system_prompt}] + history[:-1] + [
|
| 121 |
{"role": "user", "content": contenido_usuario}
|
| 122 |
]
|
| 123 |
|
| 124 |
+
# Modelo ligero para Spaces gratis
|
| 125 |
resp = client.chat.completions.create(
|
| 126 |
model="gpt-4o-mini",
|
| 127 |
messages=mensajes,
|
| 128 |
temperature=0.3,
|
| 129 |
)
|
| 130 |
+
bot = resp.choices[0].message.content
|
| 131 |
+
history.append({"role": "assistant", "content": bot})
|
|
|
|
|
|
|
| 132 |
return "", history
|
| 133 |
|
| 134 |
+
except _AuthenticationError:
|
| 135 |
+
history.append({"role": "assistant", "content":
|
| 136 |
+
"⚠️ Error de autenticación con OpenAI.\nRevisa **OPENAI_API_KEY** en Settings → Variables."})
|
| 137 |
+
return "", history
|
| 138 |
+
except _APIConnectionError:
|
| 139 |
+
history.append({"role": "assistant", "content":
|
| 140 |
+
"⚠️ Error de conexión saliente.\nActiva **Allow internet access** en Settings → Runtime/Networking."})
|
| 141 |
+
return "", history
|
| 142 |
+
except _RateLimitError:
|
| 143 |
+
history.append({"role": "assistant", "content":
|
| 144 |
+
"⚠️ Límite/ cuota de OpenAI alcanzado. Intenta más tarde o cambia de modelo."})
|
| 145 |
+
return "", history
|
| 146 |
+
except _APITimeoutError:
|
| 147 |
+
history.append({"role": "assistant", "content":
|
| 148 |
+
"⚠️ La solicitud a OpenAI excedió el tiempo de espera. Intenta de nuevo."})
|
| 149 |
+
return "", history
|
| 150 |
+
except _APIStatusError as e:
|
| 151 |
+
history.append({"role": "assistant", "content": f"⚠️ Error de API: {e}"} )
|
| 152 |
+
return "", history
|
| 153 |
except Exception as e:
|
| 154 |
+
history.append({"role": "assistant", "content": f"⚠️ Error inesperado: {e}"} )
|
| 155 |
return "", history
|
| 156 |
|
| 157 |
def limpiar_chat():
|
| 158 |
+
return [] # Chatbot(type="messages") espera una lista de dicts
|
| 159 |
|
| 160 |
# ------------------------------------------------------------
|
| 161 |
+
# UI GRADIO
|
| 162 |
# ------------------------------------------------------------
|
| 163 |
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue")) as demo:
|
| 164 |
gr.HTML("""
|
|
|
|
| 168 |
<p style="font-size:15px;">Basado en GPT-4o y los valores del IIA y Gentera</p>
|
| 169 |
</div>
|
| 170 |
""")
|
| 171 |
+
chat = gr.Chatbot(
|
| 172 |
+
label="Chat Asistente Auditoría",
|
| 173 |
+
type="messages",
|
| 174 |
+
value=[{"role": "assistant", "content": "¡Hola! Soy tu Asistente IA de Auditoría Interna. ¿En qué te ayudo hoy?"}]
|
| 175 |
+
)
|
| 176 |
msg = gr.Textbox(placeholder="Escribe tu consulta aquí...", label="Tu mensaje")
|
| 177 |
clear = gr.Button("🧹 Limpiar chat")
|
| 178 |
|