Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files- rag_books_mcp/__init__.py +10 -0
- rag_books_mcp/app.py +158 -0
- rag_books_mcp/ingest.py +244 -0
- rag_books_mcp/server.py +75 -0
- rag_books_mcp/tools.py +412 -0
rag_books_mcp/__init__.py
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# RAG Books MCP Server v2
|
| 2 |
+
"""MCP Server v2 que expone herramientas RAG sobre ESL e ISLP.
|
| 3 |
+
|
| 4 |
+
Diferencia clave con v1: la base vectorial ChromaDB no se empaqueta junto al
|
| 5 |
+
código. Se publica como dataset en HF Hub (`gusdelact/rag-esl-islp-chromadb`
|
| 6 |
+
por default) y este server hace `snapshot_download` la primera vez que se
|
| 7 |
+
necesita. Ver `tools.py` para los detalles.
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
__version__ = "2.0.0"
|
rag_books_mcp/app.py
ADDED
|
@@ -0,0 +1,158 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Gradio app v2 que expone las 4 tools como MCP Server (streamable HTTP).
|
| 3 |
+
|
| 4 |
+
Diferencia con v1: la base ChromaDB se descarga del dataset HF Hub al primer
|
| 5 |
+
uso. La carga del modelo de embeddings + descarga del snapshot se hace lazy
|
| 6 |
+
en la primera tool call, no al arrancar el Space.
|
| 7 |
+
|
| 8 |
+
Local:
|
| 9 |
+
uv run python -m rag_books_mcp.app
|
| 10 |
+
|
| 11 |
+
HF Spaces:
|
| 12 |
+
Ver `deploy_to_hf_space.py`.
|
| 13 |
+
"""
|
| 14 |
+
|
| 15 |
+
from __future__ import annotations
|
| 16 |
+
|
| 17 |
+
import gradio as gr
|
| 18 |
+
|
| 19 |
+
from rag_books_mcp.tools import (
|
| 20 |
+
cite_foundation,
|
| 21 |
+
get_section,
|
| 22 |
+
list_available_topics,
|
| 23 |
+
search_theory,
|
| 24 |
+
)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
def _build_search_tab() -> gr.Interface:
|
| 28 |
+
return gr.Interface(
|
| 29 |
+
fn=search_theory,
|
| 30 |
+
inputs=[
|
| 31 |
+
gr.Textbox(
|
| 32 |
+
label="query",
|
| 33 |
+
value="bias-variance tradeoff",
|
| 34 |
+
placeholder="Consulta en lenguaje natural",
|
| 35 |
+
),
|
| 36 |
+
gr.Radio(choices=["both", "esl", "islp"], value="both", label="book"),
|
| 37 |
+
gr.Slider(minimum=1, maximum=10, step=1, value=5, label="top_k"),
|
| 38 |
+
],
|
| 39 |
+
outputs=gr.Markdown(label="Resultados"),
|
| 40 |
+
title="🔎 search_theory",
|
| 41 |
+
description=(
|
| 42 |
+
"Búsqueda semántica en ESL e ISLP. Devuelve los fragmentos más "
|
| 43 |
+
"relevantes ordenados por similitud."
|
| 44 |
+
),
|
| 45 |
+
api_name="search_theory",
|
| 46 |
+
)
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def _build_get_section_tab() -> gr.Interface:
|
| 50 |
+
return gr.Interface(
|
| 51 |
+
fn=get_section,
|
| 52 |
+
inputs=[
|
| 53 |
+
gr.Radio(choices=["esl", "islp"], value="islp", label="book"),
|
| 54 |
+
gr.Textbox(
|
| 55 |
+
label="chapter",
|
| 56 |
+
value="8 Tree-Based Methods",
|
| 57 |
+
placeholder="Nombre del capítulo (búsqueda parcial soportada)",
|
| 58 |
+
),
|
| 59 |
+
gr.Textbox(
|
| 60 |
+
label="section",
|
| 61 |
+
value="",
|
| 62 |
+
placeholder="(Opcional) Nombre de la sección",
|
| 63 |
+
),
|
| 64 |
+
gr.Slider(minimum=1, maximum=15, step=1, value=5, label="max_chunks"),
|
| 65 |
+
],
|
| 66 |
+
outputs=gr.Markdown(label="Sección"),
|
| 67 |
+
title="📑 get_section",
|
| 68 |
+
description=(
|
| 69 |
+
"Recupera una sección específica de ESL o ISLP. Si no se encuentra "
|
| 70 |
+
"por metadata, hace fallback a búsqueda semántica."
|
| 71 |
+
),
|
| 72 |
+
api_name="get_section",
|
| 73 |
+
)
|
| 74 |
+
|
| 75 |
+
|
| 76 |
+
def _build_cite_tab() -> gr.Interface:
|
| 77 |
+
return gr.Interface(
|
| 78 |
+
fn=cite_foundation,
|
| 79 |
+
inputs=[
|
| 80 |
+
gr.Textbox(
|
| 81 |
+
label="topic",
|
| 82 |
+
value="ridge regression",
|
| 83 |
+
placeholder="Tema a fundamentar",
|
| 84 |
+
),
|
| 85 |
+
gr.Radio(
|
| 86 |
+
choices=["brief", "medium", "deep"],
|
| 87 |
+
value="medium",
|
| 88 |
+
label="detail_level",
|
| 89 |
+
),
|
| 90 |
+
],
|
| 91 |
+
outputs=gr.Markdown(label="Fundamentación"),
|
| 92 |
+
title="📚 cite_foundation",
|
| 93 |
+
description=(
|
| 94 |
+
"Fundamentación teórica que cita ambos libros: ISLP (intuitivo) y "
|
| 95 |
+
"ESL (riguroso)."
|
| 96 |
+
),
|
| 97 |
+
api_name="cite_foundation",
|
| 98 |
+
)
|
| 99 |
+
|
| 100 |
+
|
| 101 |
+
def _build_list_topics_tab() -> gr.Interface:
|
| 102 |
+
return gr.Interface(
|
| 103 |
+
fn=list_available_topics,
|
| 104 |
+
inputs=[],
|
| 105 |
+
outputs=gr.Markdown(label="Contenido indexado"),
|
| 106 |
+
title="🗂️ list_available_topics",
|
| 107 |
+
description="Lista los capítulos y secciones indexados en ChromaDB.",
|
| 108 |
+
api_name="list_available_topics",
|
| 109 |
+
)
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
def build_demo() -> gr.Blocks:
|
| 113 |
+
"""Construye la UI tabulada del MCP Server v2."""
|
| 114 |
+
with gr.Blocks(title="rag-books-mcp v2 · ESL + ISLP") as demo:
|
| 115 |
+
gr.Markdown(
|
| 116 |
+
"""
|
| 117 |
+
# 📖 RAG Books MCP v2 — ESL + ISLP
|
| 118 |
+
|
| 119 |
+
Servidor MCP que expone búsqueda semántica sobre dos libros de
|
| 120 |
+
referencia de Statistical Learning:
|
| 121 |
+
|
| 122 |
+
- **ESL** — *The Elements of Statistical Learning* (Hastie, Tibshirani, Friedman)
|
| 123 |
+
- **ISLP** — *An Introduction to Statistical Learning with Python* (James, Witten, Hastie, Tibshirani)
|
| 124 |
+
|
| 125 |
+
**v2 vs v1:** la base ChromaDB se carga desde el dataset HF
|
| 126 |
+
`gusdelact/rag-esl-islp-chromadb` en lugar de empaquetarla con el
|
| 127 |
+
código. Permite versionar el índice independientemente y reusarlo
|
| 128 |
+
desde otros clientes.
|
| 129 |
+
|
| 130 |
+
**Endpoint MCP:** `/gradio_api/mcp/` (streamable HTTP).
|
| 131 |
+
**Embeddings:** `sentence-transformers/all-MiniLM-L6-v2` (local, sin API key).
|
| 132 |
+
**Vector store:** ChromaDB con 1977 chunks (1093 ESL + 884 ISLP).
|
| 133 |
+
|
| 134 |
+
La primera tool call descarga el dataset (~40 MB). Las siguientes
|
| 135 |
+
son cache hit.
|
| 136 |
+
"""
|
| 137 |
+
)
|
| 138 |
+
|
| 139 |
+
gr.TabbedInterface(
|
| 140 |
+
interface_list=[
|
| 141 |
+
_build_search_tab(),
|
| 142 |
+
_build_cite_tab(),
|
| 143 |
+
_build_get_section_tab(),
|
| 144 |
+
_build_list_topics_tab(),
|
| 145 |
+
],
|
| 146 |
+
tab_names=["search_theory", "cite_foundation", "get_section", "list_available_topics"],
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
return demo
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
def main() -> None:
|
| 153 |
+
demo = build_demo()
|
| 154 |
+
demo.launch(mcp_server=True, server_name="0.0.0.0")
|
| 155 |
+
|
| 156 |
+
|
| 157 |
+
if __name__ == "__main__":
|
| 158 |
+
main()
|
rag_books_mcp/ingest.py
ADDED
|
@@ -0,0 +1,244 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Script de ingesta: vectoriza los capítulos de ESL e ISLP en ChromaDB.
|
| 3 |
+
|
| 4 |
+
Uso:
|
| 5 |
+
python -m rag_books_mcp.ingest --books-dir ../ebook
|
| 6 |
+
|
| 7 |
+
Esto crea/actualiza la base vectorial en ./chroma_db/
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
import os
|
| 11 |
+
import re
|
| 12 |
+
import argparse
|
| 13 |
+
from pathlib import Path
|
| 14 |
+
|
| 15 |
+
import chromadb
|
| 16 |
+
from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
# --- Configuración ---
|
| 20 |
+
EMBEDDING_MODEL = "all-MiniLM-L6-v2"
|
| 21 |
+
CHUNK_SIZE = 600 # tokens aprox (caracteres / 4)
|
| 22 |
+
CHUNK_OVERLAP = 100
|
| 23 |
+
CHROMA_DIR = Path(__file__).parent.parent / "chroma_db"
|
| 24 |
+
|
| 25 |
+
BOOKS_CONFIG = {
|
| 26 |
+
"esl": {
|
| 27 |
+
"dir_name": "capitulos_TheElementsOfStatisticalLearning",
|
| 28 |
+
"collection": "esl_chapters",
|
| 29 |
+
"full_name": "The Elements of Statistical Learning (Hastie, Tibshirani, Friedman)",
|
| 30 |
+
},
|
| 31 |
+
"islp": {
|
| 32 |
+
"dir_name": "capitulos_islp",
|
| 33 |
+
"collection": "islp_chapters",
|
| 34 |
+
"full_name": "An Introduction to Statistical Learning with Python (James, Witten, Hastie, Tibshirani)",
|
| 35 |
+
},
|
| 36 |
+
}
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def extract_chapter_info(filename: str) -> dict:
|
| 40 |
+
"""Extrae número de archivo y nombre del capítulo del filename."""
|
| 41 |
+
# Formato: 04_3_Linear_Methods_for_Regression.md
|
| 42 |
+
stem = Path(filename).stem
|
| 43 |
+
parts = stem.split("_", 1)
|
| 44 |
+
file_order = parts[0] if parts else "00"
|
| 45 |
+
chapter_title = parts[1].replace("_", " ") if len(parts) > 1 else stem
|
| 46 |
+
return {"file_order": file_order, "chapter_title": chapter_title}
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def split_by_sections(text: str, chapter_title: str) -> list[dict]:
|
| 50 |
+
"""
|
| 51 |
+
Divide el texto en secciones usando headers markdown (# y ##).
|
| 52 |
+
Cada sección se subdivide en chunks si es muy larga.
|
| 53 |
+
"""
|
| 54 |
+
# Patrón para detectar headers de nivel 1-3
|
| 55 |
+
header_pattern = re.compile(r"^(#{1,3})\s+(.+)$", re.MULTILINE)
|
| 56 |
+
|
| 57 |
+
sections = []
|
| 58 |
+
matches = list(header_pattern.finditer(text))
|
| 59 |
+
|
| 60 |
+
if not matches:
|
| 61 |
+
# Sin headers, tratar todo como una sección
|
| 62 |
+
sections.append({"title": chapter_title, "level": 1, "content": text.strip()})
|
| 63 |
+
else:
|
| 64 |
+
# Texto antes del primer header
|
| 65 |
+
pre_text = text[: matches[0].start()].strip()
|
| 66 |
+
if pre_text and len(pre_text) > 50:
|
| 67 |
+
sections.append({"title": chapter_title, "level": 1, "content": pre_text})
|
| 68 |
+
|
| 69 |
+
for i, match in enumerate(matches):
|
| 70 |
+
level = len(match.group(1))
|
| 71 |
+
title = match.group(2).strip()
|
| 72 |
+
start = match.end()
|
| 73 |
+
end = matches[i + 1].start() if i + 1 < len(matches) else len(text)
|
| 74 |
+
content = text[start:end].strip()
|
| 75 |
+
|
| 76 |
+
if content and len(content) > 30:
|
| 77 |
+
sections.append({"title": title, "level": level, "content": content})
|
| 78 |
+
|
| 79 |
+
return sections
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
def chunk_text(text: str, chunk_size: int = CHUNK_SIZE, overlap: int = CHUNK_OVERLAP) -> list[str]:
|
| 83 |
+
"""
|
| 84 |
+
Divide texto en chunks por caracteres con overlap.
|
| 85 |
+
Intenta cortar en saltos de línea o puntos para no romper oraciones.
|
| 86 |
+
"""
|
| 87 |
+
# Convertir chunk_size de tokens aprox a caracteres (1 token ≈ 4 chars)
|
| 88 |
+
char_size = chunk_size * 4
|
| 89 |
+
char_overlap = overlap * 4
|
| 90 |
+
|
| 91 |
+
if len(text) <= char_size:
|
| 92 |
+
return [text]
|
| 93 |
+
|
| 94 |
+
chunks = []
|
| 95 |
+
start = 0
|
| 96 |
+
|
| 97 |
+
while start < len(text):
|
| 98 |
+
end = start + char_size
|
| 99 |
+
|
| 100 |
+
if end < len(text):
|
| 101 |
+
# Buscar un buen punto de corte (párrafo o punto)
|
| 102 |
+
# Primero intentar doble newline (párrafo)
|
| 103 |
+
cut_point = text.rfind("\n\n", start + char_size // 2, end)
|
| 104 |
+
if cut_point == -1:
|
| 105 |
+
# Intentar punto seguido de espacio
|
| 106 |
+
cut_point = text.rfind(". ", start + char_size // 2, end)
|
| 107 |
+
if cut_point != -1:
|
| 108 |
+
cut_point += 1 # incluir el punto
|
| 109 |
+
if cut_point == -1:
|
| 110 |
+
# Intentar newline simple
|
| 111 |
+
cut_point = text.rfind("\n", start + char_size // 2, end)
|
| 112 |
+
if cut_point == -1:
|
| 113 |
+
cut_point = end
|
| 114 |
+
|
| 115 |
+
end = cut_point
|
| 116 |
+
|
| 117 |
+
chunk = text[start:end].strip()
|
| 118 |
+
if chunk:
|
| 119 |
+
chunks.append(chunk)
|
| 120 |
+
|
| 121 |
+
start = end - char_overlap
|
| 122 |
+
if start >= len(text):
|
| 123 |
+
break
|
| 124 |
+
|
| 125 |
+
return chunks
|
| 126 |
+
|
| 127 |
+
|
| 128 |
+
def clean_text(text: str) -> str:
|
| 129 |
+
"""Limpia artefactos de la extracción PDF."""
|
| 130 |
+
# Eliminar marcadores de página
|
| 131 |
+
text = re.sub(r"---\s*Página\s*\d+\s*---", "", text)
|
| 132 |
+
# Eliminar líneas con solo números (números de página sueltos)
|
| 133 |
+
text = re.sub(r"^\d+\s*$", "", text, flags=re.MULTILINE)
|
| 134 |
+
# Reducir múltiples líneas vacías
|
| 135 |
+
text = re.sub(r"\n{4,}", "\n\n\n", text)
|
| 136 |
+
# Eliminar copyright notices
|
| 137 |
+
text = re.sub(r"©.*?(?:\n|$)", "", text)
|
| 138 |
+
return text.strip()
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
def ingest_book(books_dir: Path, book_key: str, client: chromadb.ClientAPI, embedding_fn):
|
| 142 |
+
"""Ingesta un libro completo en ChromaDB."""
|
| 143 |
+
config = BOOKS_CONFIG[book_key]
|
| 144 |
+
chapters_dir = books_dir / config["dir_name"]
|
| 145 |
+
|
| 146 |
+
if not chapters_dir.exists():
|
| 147 |
+
print(f" ⚠️ Directorio no encontrado: {chapters_dir}")
|
| 148 |
+
return 0
|
| 149 |
+
|
| 150 |
+
# Crear o obtener colección (reset si existe)
|
| 151 |
+
try:
|
| 152 |
+
client.delete_collection(config["collection"])
|
| 153 |
+
except Exception:
|
| 154 |
+
pass
|
| 155 |
+
|
| 156 |
+
collection = client.get_or_create_collection(
|
| 157 |
+
name=config["collection"],
|
| 158 |
+
embedding_function=embedding_fn,
|
| 159 |
+
metadata={"hnsw:space": "cosine"},
|
| 160 |
+
)
|
| 161 |
+
|
| 162 |
+
total_chunks = 0
|
| 163 |
+
md_files = sorted(chapters_dir.glob("*.md"))
|
| 164 |
+
|
| 165 |
+
print(f"\n 📚 {config['full_name']}")
|
| 166 |
+
print(f" Archivos encontrados: {len(md_files)}")
|
| 167 |
+
|
| 168 |
+
for md_file in md_files:
|
| 169 |
+
chapter_info = extract_chapter_info(md_file.name)
|
| 170 |
+
raw_text = md_file.read_text(encoding="utf-8")
|
| 171 |
+
text = clean_text(raw_text)
|
| 172 |
+
|
| 173 |
+
if len(text) < 100:
|
| 174 |
+
continue
|
| 175 |
+
|
| 176 |
+
# Dividir en secciones
|
| 177 |
+
sections = split_by_sections(text, chapter_info["chapter_title"])
|
| 178 |
+
|
| 179 |
+
for section in sections:
|
| 180 |
+
# Dividir secciones largas en chunks
|
| 181 |
+
chunks = chunk_text(section["content"])
|
| 182 |
+
|
| 183 |
+
for i, chunk in enumerate(chunks):
|
| 184 |
+
chunk_id = f"{book_key}_{chapter_info['file_order']}_{section['title'][:30]}_{i}"
|
| 185 |
+
# Sanitizar ID
|
| 186 |
+
chunk_id = re.sub(r"[^a-zA-Z0-9_-]", "_", chunk_id)
|
| 187 |
+
|
| 188 |
+
metadata = {
|
| 189 |
+
"book": book_key,
|
| 190 |
+
"book_full_name": config["full_name"],
|
| 191 |
+
"chapter": chapter_info["chapter_title"],
|
| 192 |
+
"section": section["title"],
|
| 193 |
+
"section_level": section["level"],
|
| 194 |
+
"chunk_index": i,
|
| 195 |
+
"total_chunks_in_section": len(chunks),
|
| 196 |
+
"file": md_file.name,
|
| 197 |
+
}
|
| 198 |
+
|
| 199 |
+
collection.add(
|
| 200 |
+
ids=[chunk_id],
|
| 201 |
+
documents=[chunk],
|
| 202 |
+
metadatas=[metadata],
|
| 203 |
+
)
|
| 204 |
+
total_chunks += 1
|
| 205 |
+
|
| 206 |
+
print(f" ✓ {md_file.name} → {len(sections)} secciones")
|
| 207 |
+
|
| 208 |
+
print(f" Total chunks: {total_chunks}")
|
| 209 |
+
return total_chunks
|
| 210 |
+
|
| 211 |
+
|
| 212 |
+
def main():
|
| 213 |
+
parser = argparse.ArgumentParser(description="Ingesta de libros ESL/ISLP en ChromaDB")
|
| 214 |
+
parser.add_argument(
|
| 215 |
+
"--books-dir",
|
| 216 |
+
type=Path,
|
| 217 |
+
default=Path(__file__).parent.parent.parent / "ebook",
|
| 218 |
+
help="Directorio raíz con las carpetas de capítulos",
|
| 219 |
+
)
|
| 220 |
+
parser.add_argument(
|
| 221 |
+
"--chroma-dir",
|
| 222 |
+
type=Path,
|
| 223 |
+
default=CHROMA_DIR,
|
| 224 |
+
help="Directorio para la base de datos ChromaDB",
|
| 225 |
+
)
|
| 226 |
+
args = parser.parse_args()
|
| 227 |
+
|
| 228 |
+
print("🔧 Inicializando embedding model...")
|
| 229 |
+
embedding_fn = SentenceTransformerEmbeddingFunction(model_name=EMBEDDING_MODEL)
|
| 230 |
+
|
| 231 |
+
print(f"🗄️ ChromaDB persistente en: {args.chroma_dir}")
|
| 232 |
+
client = chromadb.PersistentClient(path=str(args.chroma_dir))
|
| 233 |
+
|
| 234 |
+
print("\n📖 Iniciando ingesta de libros...")
|
| 235 |
+
total = 0
|
| 236 |
+
for book_key in BOOKS_CONFIG:
|
| 237 |
+
total += ingest_book(args.books_dir, book_key, client, embedding_fn)
|
| 238 |
+
|
| 239 |
+
print(f"\n✅ Ingesta completada. Total de chunks vectorizados: {total}")
|
| 240 |
+
print(f" Base de datos en: {args.chroma_dir}")
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
if __name__ == "__main__":
|
| 244 |
+
main()
|
rag_books_mcp/server.py
ADDED
|
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
MCP Server v2 (transporte stdio) — RAG sobre ESL e ISLP.
|
| 3 |
+
|
| 4 |
+
Diferencia con v1: la base ChromaDB se obtiene de un dataset HF Hub
|
| 5 |
+
(ver `rag_books_mcp.tools` para la resolución de la ruta).
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
from mcp.server.fastmcp import FastMCP
|
| 9 |
+
|
| 10 |
+
from rag_books_mcp.tools import (
|
| 11 |
+
cite_foundation as _cite_foundation,
|
| 12 |
+
get_section as _get_section,
|
| 13 |
+
list_available_topics as _list_available_topics,
|
| 14 |
+
search_theory as _search_theory,
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
mcp = FastMCP(
|
| 19 |
+
"rag-books-mcp-v2",
|
| 20 |
+
instructions=(
|
| 21 |
+
"RAG sobre los libros ESL e ISLP. v2: base vectorial ChromaDB cargada "
|
| 22 |
+
"desde un dataset publicado en HF Hub (separación código/datos)."
|
| 23 |
+
),
|
| 24 |
+
)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
@mcp.tool()
|
| 28 |
+
def search_theory(query: str, book: str = "both", top_k: int = 5) -> str:
|
| 29 |
+
"""Busca fragmentos relevantes en ESL/ISLP usando búsqueda semántica.
|
| 30 |
+
|
| 31 |
+
Args:
|
| 32 |
+
query: Consulta en lenguaje natural (ej: "bias-variance tradeoff").
|
| 33 |
+
book: "esl", "islp" o "both" (default: "both").
|
| 34 |
+
top_k: Número de resultados (1-10, default: 5).
|
| 35 |
+
"""
|
| 36 |
+
return _search_theory(query=query, book=book, top_k=top_k)
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
@mcp.tool()
|
| 40 |
+
def get_section(book: str, chapter: str, section: str = "", max_chunks: int = 5) -> str:
|
| 41 |
+
"""Recupera una sección específica de ESL o ISLP por referencia exacta.
|
| 42 |
+
|
| 43 |
+
Args:
|
| 44 |
+
book: "esl" o "islp".
|
| 45 |
+
chapter: Nombre del capítulo (búsqueda parcial soportada).
|
| 46 |
+
section: Nombre de la sección dentro del capítulo (opcional).
|
| 47 |
+
max_chunks: Máximo de chunks a devolver (default: 5).
|
| 48 |
+
"""
|
| 49 |
+
return _get_section(book=book, chapter=chapter, section=section, max_chunks=max_chunks)
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
@mcp.tool()
|
| 53 |
+
def cite_foundation(topic: str, detail_level: str = "medium") -> str:
|
| 54 |
+
"""Fundamentación teórica de un tema citando ambos libros (ESL + ISLP).
|
| 55 |
+
|
| 56 |
+
Args:
|
| 57 |
+
topic: Tema a fundamentar (ej: "ridge regression", "bagging").
|
| 58 |
+
detail_level: "brief", "medium" (default) o "deep".
|
| 59 |
+
"""
|
| 60 |
+
return _cite_foundation(topic=topic, detail_level=detail_level)
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
@mcp.tool()
|
| 64 |
+
def list_available_topics() -> str:
|
| 65 |
+
"""Lista los capítulos y temas indexados en la base de conocimiento."""
|
| 66 |
+
return _list_available_topics()
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def main():
|
| 70 |
+
"""Punto de entrada del MCP server (stdio)."""
|
| 71 |
+
mcp.run(transport="stdio")
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
if __name__ == "__main__":
|
| 75 |
+
main()
|
rag_books_mcp/tools.py
ADDED
|
@@ -0,0 +1,412 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Lógica de las 4 tools de RAG sobre ESL e ISLP (v2).
|
| 3 |
+
|
| 4 |
+
Diferencia con v1: la base ChromaDB se obtiene de un dataset publicado en
|
| 5 |
+
HF Hub vía `snapshot_download`. La primera invocación tarda lo que tarde
|
| 6 |
+
la descarga (~40 MB); las siguientes son cache hit.
|
| 7 |
+
|
| 8 |
+
Variables de entorno:
|
| 9 |
+
- RAG_CHROMA_DIR Si está set y apunta a una carpeta existente, se usa
|
| 10 |
+
en lugar del dataset (útil para dev local con índice
|
| 11 |
+
recién regenerado por `ingest.py`).
|
| 12 |
+
- RAG_CHROMA_DATASET Repo del dataset HF a descargar.
|
| 13 |
+
Default: gusdelact/rag-esl-islp-chromadb
|
| 14 |
+
- RAG_CHROMA_REVISION Revision (branch/tag/commit) del dataset.
|
| 15 |
+
Default: main
|
| 16 |
+
- RAG_CHROMA_CACHE_DIR Directorio cache para el snapshot_download.
|
| 17 |
+
Default: ~/.cache/rag-books-mcp/chroma_db (o /data/chroma_db
|
| 18 |
+
si existe /data, como en HF Spaces con persistent storage).
|
| 19 |
+
- HF_TOKEN Solo si el dataset es privado.
|
| 20 |
+
"""
|
| 21 |
+
|
| 22 |
+
from __future__ import annotations
|
| 23 |
+
|
| 24 |
+
import os
|
| 25 |
+
import sys
|
| 26 |
+
from pathlib import Path
|
| 27 |
+
from typing import Optional
|
| 28 |
+
|
| 29 |
+
import chromadb
|
| 30 |
+
from chromadb.utils.embedding_functions import SentenceTransformerEmbeddingFunction
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
# --- Configuración ---
|
| 34 |
+
EMBEDDING_MODEL = "all-MiniLM-L6-v2"
|
| 35 |
+
|
| 36 |
+
DEFAULT_DATASET = "gusdelact/rag-esl-islp-chromadb"
|
| 37 |
+
DEFAULT_REVISION = "main"
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
def _resolve_cache_dir() -> Path:
|
| 41 |
+
"""Decide dónde guardar el snapshot del dataset.
|
| 42 |
+
|
| 43 |
+
Prioridad:
|
| 44 |
+
1. RAG_CHROMA_CACHE_DIR si está set.
|
| 45 |
+
2. /data/chroma_db si existe /data (HF Spaces con persistent storage).
|
| 46 |
+
3. ~/.cache/rag-books-mcp/chroma_db.
|
| 47 |
+
"""
|
| 48 |
+
explicit = os.environ.get("RAG_CHROMA_CACHE_DIR")
|
| 49 |
+
if explicit:
|
| 50 |
+
return Path(explicit)
|
| 51 |
+
if Path("/data").is_dir():
|
| 52 |
+
return Path("/data/chroma_db")
|
| 53 |
+
return Path.home() / ".cache" / "rag-books-mcp" / "chroma_db"
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
# Singletons por proceso
|
| 57 |
+
_client: Optional[chromadb.ClientAPI] = None
|
| 58 |
+
_embedding_fn = None
|
| 59 |
+
_chroma_path_resolved: Optional[str] = None
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def _resolve_chroma_path() -> str:
|
| 63 |
+
"""Resuelve la ruta a usar como ChromaDB persistente.
|
| 64 |
+
|
| 65 |
+
Si `RAG_CHROMA_DIR` apunta a una carpeta existente, la usa tal cual.
|
| 66 |
+
En caso contrario, baja `RAG_CHROMA_DATASET@RAG_CHROMA_REVISION` desde HF
|
| 67 |
+
Hub y devuelve la ruta al snapshot.
|
| 68 |
+
"""
|
| 69 |
+
global _chroma_path_resolved
|
| 70 |
+
if _chroma_path_resolved is not None:
|
| 71 |
+
return _chroma_path_resolved
|
| 72 |
+
|
| 73 |
+
override = os.environ.get("RAG_CHROMA_DIR")
|
| 74 |
+
if override and Path(override).is_dir():
|
| 75 |
+
print(f"[rag-books-mcp v2] Using local RAG_CHROMA_DIR: {override}", file=sys.stderr)
|
| 76 |
+
_chroma_path_resolved = override
|
| 77 |
+
return override
|
| 78 |
+
|
| 79 |
+
repo_id = os.environ.get("RAG_CHROMA_DATASET", DEFAULT_DATASET)
|
| 80 |
+
revision = os.environ.get("RAG_CHROMA_REVISION", DEFAULT_REVISION)
|
| 81 |
+
cache_dir = _resolve_cache_dir()
|
| 82 |
+
cache_dir.mkdir(parents=True, exist_ok=True)
|
| 83 |
+
|
| 84 |
+
print(
|
| 85 |
+
f"[rag-books-mcp v2] Downloading ChromaDB from HF dataset "
|
| 86 |
+
f"{repo_id}@{revision} into {cache_dir} ...",
|
| 87 |
+
file=sys.stderr,
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
# Import perezoso para no pagar el costo si está cacheado vía RAG_CHROMA_DIR.
|
| 91 |
+
from huggingface_hub import snapshot_download
|
| 92 |
+
|
| 93 |
+
snapshot = snapshot_download(
|
| 94 |
+
repo_id=repo_id,
|
| 95 |
+
repo_type="dataset",
|
| 96 |
+
revision=revision,
|
| 97 |
+
cache_dir=str(cache_dir),
|
| 98 |
+
token=os.environ.get("HF_TOKEN"), # solo si es privado
|
| 99 |
+
)
|
| 100 |
+
|
| 101 |
+
print(f"[rag-books-mcp v2] ChromaDB ready at {snapshot}", file=sys.stderr)
|
| 102 |
+
_chroma_path_resolved = snapshot
|
| 103 |
+
return snapshot
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def get_client() -> chromadb.ClientAPI:
|
| 107 |
+
"""Cliente ChromaDB persistente (singleton)."""
|
| 108 |
+
global _client
|
| 109 |
+
if _client is None:
|
| 110 |
+
_client = chromadb.PersistentClient(path=_resolve_chroma_path())
|
| 111 |
+
return _client
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
def get_embedding_fn():
|
| 115 |
+
"""Función de embedding `sentence-transformers/all-MiniLM-L6-v2` (singleton)."""
|
| 116 |
+
global _embedding_fn
|
| 117 |
+
if _embedding_fn is None:
|
| 118 |
+
_embedding_fn = SentenceTransformerEmbeddingFunction(model_name=EMBEDDING_MODEL)
|
| 119 |
+
return _embedding_fn
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
def get_collection(name: str):
|
| 123 |
+
"""Obtiene una colección de ChromaDB por nombre."""
|
| 124 |
+
return get_client().get_collection(name=name, embedding_function=get_embedding_fn())
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
# --- Tools (idénticas en comportamiento a v1) ---
|
| 128 |
+
|
| 129 |
+
def search_theory(
|
| 130 |
+
query: str,
|
| 131 |
+
book: str = "both",
|
| 132 |
+
top_k: int = 5,
|
| 133 |
+
) -> str:
|
| 134 |
+
"""
|
| 135 |
+
Busca fragmentos relevantes en los libros ESL e ISLP usando búsqueda semántica.
|
| 136 |
+
|
| 137 |
+
Args:
|
| 138 |
+
query (str): Consulta en lenguaje natural (ej: "bias-variance tradeoff",
|
| 139 |
+
"regularización L1 vs L2", "random forest out-of-bag error").
|
| 140 |
+
book (str): Libro donde buscar. Opciones: "esl", "islp", "both" (default: "both").
|
| 141 |
+
top_k (int): Número de resultados a devolver (default: 5, máximo: 10).
|
| 142 |
+
|
| 143 |
+
Returns:
|
| 144 |
+
str: Fragmentos relevantes con metadata (libro, capítulo, sección, similitud).
|
| 145 |
+
"""
|
| 146 |
+
top_k = min(max(int(top_k), 1), 10)
|
| 147 |
+
|
| 148 |
+
collections_to_search = []
|
| 149 |
+
if book in ("esl", "both"):
|
| 150 |
+
try:
|
| 151 |
+
collections_to_search.append(("ESL", get_collection("esl_chapters")))
|
| 152 |
+
except Exception:
|
| 153 |
+
pass
|
| 154 |
+
if book in ("islp", "both"):
|
| 155 |
+
try:
|
| 156 |
+
collections_to_search.append(("ISLP", get_collection("islp_chapters")))
|
| 157 |
+
except Exception:
|
| 158 |
+
pass
|
| 159 |
+
|
| 160 |
+
if not collections_to_search:
|
| 161 |
+
return (
|
| 162 |
+
"❌ No se encontraron colecciones. Verifica que el dataset HF "
|
| 163 |
+
"esté disponible o ejecuta la ingesta local."
|
| 164 |
+
)
|
| 165 |
+
|
| 166 |
+
results = []
|
| 167 |
+
for book_label, collection in collections_to_search:
|
| 168 |
+
res = collection.query(query_texts=[query], n_results=top_k)
|
| 169 |
+
|
| 170 |
+
if res["documents"] and res["documents"][0]:
|
| 171 |
+
for doc, meta, dist in zip(
|
| 172 |
+
res["documents"][0], res["metadatas"][0], res["distances"][0]
|
| 173 |
+
):
|
| 174 |
+
similarity = 1 - dist
|
| 175 |
+
results.append({
|
| 176 |
+
"book": book_label,
|
| 177 |
+
"chapter": meta.get("chapter", ""),
|
| 178 |
+
"section": meta.get("section", ""),
|
| 179 |
+
"similarity": similarity,
|
| 180 |
+
"content": doc,
|
| 181 |
+
})
|
| 182 |
+
|
| 183 |
+
results.sort(key=lambda x: x["similarity"], reverse=True)
|
| 184 |
+
results = results[:top_k]
|
| 185 |
+
|
| 186 |
+
if not results:
|
| 187 |
+
return f"No se encontraron resultados para: '{query}'"
|
| 188 |
+
|
| 189 |
+
output_parts = [f"## Resultados para: \"{query}\"\n"]
|
| 190 |
+
for i, r in enumerate(results, 1):
|
| 191 |
+
output_parts.append(
|
| 192 |
+
f"### [{i}] {r['book']} — {r['chapter']} § {r['section']}\n"
|
| 193 |
+
f"**Similitud:** {r['similarity']:.3f}\n\n"
|
| 194 |
+
f"{r['content'][:1500]}\n\n---\n"
|
| 195 |
+
)
|
| 196 |
+
return "\n".join(output_parts)
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
def get_section(
|
| 200 |
+
book: str,
|
| 201 |
+
chapter: str,
|
| 202 |
+
section: str = "",
|
| 203 |
+
max_chunks: int = 5,
|
| 204 |
+
) -> str:
|
| 205 |
+
"""
|
| 206 |
+
Recupera una sección específica de un libro por referencia exacta.
|
| 207 |
+
|
| 208 |
+
Args:
|
| 209 |
+
book (str): Libro a consultar. Opciones: "esl" o "islp".
|
| 210 |
+
chapter (str): Nombre del capítulo (búsqueda parcial soportada).
|
| 211 |
+
section (str): Nombre de la sección dentro del capítulo (opcional).
|
| 212 |
+
max_chunks (int): Máximo de chunks a devolver (default: 5).
|
| 213 |
+
|
| 214 |
+
Returns:
|
| 215 |
+
str: Contenido de la sección con metadata.
|
| 216 |
+
"""
|
| 217 |
+
max_chunks = int(max_chunks)
|
| 218 |
+
collection_name = f"{book}_chapters"
|
| 219 |
+
try:
|
| 220 |
+
collection = get_collection(collection_name)
|
| 221 |
+
except Exception:
|
| 222 |
+
return f"❌ Colección '{collection_name}' no encontrada. Opciones: esl, islp"
|
| 223 |
+
|
| 224 |
+
try:
|
| 225 |
+
if section:
|
| 226 |
+
results = collection.get(
|
| 227 |
+
where={"$and": [
|
| 228 |
+
{"chapter": {"$contains": chapter}},
|
| 229 |
+
{"section": {"$contains": section}},
|
| 230 |
+
]},
|
| 231 |
+
limit=max_chunks,
|
| 232 |
+
)
|
| 233 |
+
else:
|
| 234 |
+
results = collection.get(
|
| 235 |
+
where={"chapter": {"$contains": chapter}},
|
| 236 |
+
limit=max_chunks,
|
| 237 |
+
)
|
| 238 |
+
except Exception:
|
| 239 |
+
search_query = f"{chapter} {section}".strip()
|
| 240 |
+
results = collection.query(query_texts=[search_query], n_results=max_chunks)
|
| 241 |
+
if results["documents"] and results["documents"][0]:
|
| 242 |
+
output_parts = [f"## {book.upper()} — {chapter}\n"]
|
| 243 |
+
for doc, meta in zip(results["documents"][0], results["metadatas"][0]):
|
| 244 |
+
output_parts.append(
|
| 245 |
+
f"### § {meta.get('section', 'N/A')}\n\n{doc}\n\n---\n"
|
| 246 |
+
)
|
| 247 |
+
return "\n".join(output_parts)
|
| 248 |
+
return f"No se encontró el capítulo '{chapter}' en {book.upper()}"
|
| 249 |
+
|
| 250 |
+
if not results["documents"]:
|
| 251 |
+
search_query = f"{chapter} {section}".strip()
|
| 252 |
+
results = collection.query(query_texts=[search_query], n_results=max_chunks)
|
| 253 |
+
if results["documents"] and results["documents"][0]:
|
| 254 |
+
output_parts = [f"## {book.upper()} — {chapter}\n"]
|
| 255 |
+
for doc, meta in zip(results["documents"][0], results["metadatas"][0]):
|
| 256 |
+
output_parts.append(
|
| 257 |
+
f"### § {meta.get('section', 'N/A')}\n\n{doc}\n\n---\n"
|
| 258 |
+
)
|
| 259 |
+
return "\n".join(output_parts)
|
| 260 |
+
return f"No se encontró el capítulo '{chapter}' en {book.upper()}"
|
| 261 |
+
|
| 262 |
+
output_parts = [f"## {book.upper()} — {chapter}"]
|
| 263 |
+
if section:
|
| 264 |
+
output_parts[0] += f" § {section}"
|
| 265 |
+
output_parts[0] += "\n"
|
| 266 |
+
|
| 267 |
+
for doc, meta in zip(results["documents"], results["metadatas"]):
|
| 268 |
+
sec_title = meta.get("section", "")
|
| 269 |
+
chunk_idx = meta.get("chunk_index", 0)
|
| 270 |
+
total = meta.get("total_chunks_in_section", 1)
|
| 271 |
+
output_parts.append(
|
| 272 |
+
f"### § {sec_title} (parte {chunk_idx + 1}/{total})\n\n{doc}\n\n---\n"
|
| 273 |
+
)
|
| 274 |
+
return "\n".join(output_parts)
|
| 275 |
+
|
| 276 |
+
|
| 277 |
+
def cite_foundation(
|
| 278 |
+
topic: str,
|
| 279 |
+
detail_level: str = "medium",
|
| 280 |
+
) -> str:
|
| 281 |
+
"""
|
| 282 |
+
Devuelve la fundamentación teórica para un tema citando ambos libros (ESL + ISLP).
|
| 283 |
+
|
| 284 |
+
Args:
|
| 285 |
+
topic (str): Tema a fundamentar (ej: "ridge regression", "bagging").
|
| 286 |
+
detail_level (str): "brief" (1-2), "medium" (3-4) o "deep" (6-8).
|
| 287 |
+
|
| 288 |
+
Returns:
|
| 289 |
+
str: Fundamentación teórica con citas, organizada de intuitivo (ISLP)
|
| 290 |
+
a riguroso (ESL).
|
| 291 |
+
"""
|
| 292 |
+
top_k_map = {"brief": 2, "medium": 4, "deep": 8}
|
| 293 |
+
top_k = top_k_map.get(detail_level, 4)
|
| 294 |
+
|
| 295 |
+
islp_results = []
|
| 296 |
+
try:
|
| 297 |
+
islp_col = get_collection("islp_chapters")
|
| 298 |
+
res = islp_col.query(query_texts=[topic], n_results=top_k)
|
| 299 |
+
if res["documents"] and res["documents"][0]:
|
| 300 |
+
for doc, meta, dist in zip(
|
| 301 |
+
res["documents"][0], res["metadatas"][0], res["distances"][0]
|
| 302 |
+
):
|
| 303 |
+
islp_results.append({
|
| 304 |
+
"content": doc,
|
| 305 |
+
"chapter": meta.get("chapter", ""),
|
| 306 |
+
"section": meta.get("section", ""),
|
| 307 |
+
"similarity": 1 - dist,
|
| 308 |
+
})
|
| 309 |
+
except Exception:
|
| 310 |
+
pass
|
| 311 |
+
|
| 312 |
+
esl_results = []
|
| 313 |
+
try:
|
| 314 |
+
esl_col = get_collection("esl_chapters")
|
| 315 |
+
res = esl_col.query(query_texts=[topic], n_results=top_k)
|
| 316 |
+
if res["documents"] and res["documents"][0]:
|
| 317 |
+
for doc, meta, dist in zip(
|
| 318 |
+
res["documents"][0], res["metadatas"][0], res["distances"][0]
|
| 319 |
+
):
|
| 320 |
+
esl_results.append({
|
| 321 |
+
"content": doc,
|
| 322 |
+
"chapter": meta.get("chapter", ""),
|
| 323 |
+
"section": meta.get("section", ""),
|
| 324 |
+
"similarity": 1 - dist,
|
| 325 |
+
})
|
| 326 |
+
except Exception:
|
| 327 |
+
pass
|
| 328 |
+
|
| 329 |
+
if not islp_results and not esl_results:
|
| 330 |
+
return (
|
| 331 |
+
f"❌ No se encontró fundamentación para '{topic}'. "
|
| 332 |
+
"Verifica que la ingesta se haya ejecutado correctamente."
|
| 333 |
+
)
|
| 334 |
+
|
| 335 |
+
output_parts = [
|
| 336 |
+
f"# Fundamentación Teórica: {topic}\n",
|
| 337 |
+
f"**Nivel de detalle:** {detail_level}\n",
|
| 338 |
+
]
|
| 339 |
+
|
| 340 |
+
if islp_results:
|
| 341 |
+
output_parts.append("\n## 📘 ISLP (Explicación Intuitiva)\n")
|
| 342 |
+
for i, r in enumerate(islp_results, 1):
|
| 343 |
+
output_parts.append(
|
| 344 |
+
f"### [{i}] Cap. {r['chapter']} § {r['section']} "
|
| 345 |
+
f"(sim: {r['similarity']:.3f})\n\n"
|
| 346 |
+
f"{r['content'][:1200]}\n\n---\n"
|
| 347 |
+
)
|
| 348 |
+
|
| 349 |
+
if esl_results:
|
| 350 |
+
output_parts.append("\n## 📗 ESL (Tratamiento Riguroso)\n")
|
| 351 |
+
for i, r in enumerate(esl_results, 1):
|
| 352 |
+
output_parts.append(
|
| 353 |
+
f"### [{i}] Cap. {r['chapter']} § {r['section']} "
|
| 354 |
+
f"(sim: {r['similarity']:.3f})\n\n"
|
| 355 |
+
f"{r['content'][:1200]}\n\n---\n"
|
| 356 |
+
)
|
| 357 |
+
|
| 358 |
+
output_parts.append("\n## 📚 Referencias\n")
|
| 359 |
+
if islp_results:
|
| 360 |
+
chapters = set(r["chapter"] for r in islp_results)
|
| 361 |
+
output_parts.append(f"- **ISLP:** {', '.join(chapters)}\n")
|
| 362 |
+
if esl_results:
|
| 363 |
+
chapters = set(r["chapter"] for r in esl_results)
|
| 364 |
+
output_parts.append(f"- **ESL:** {', '.join(chapters)}\n")
|
| 365 |
+
|
| 366 |
+
return "\n".join(output_parts)
|
| 367 |
+
|
| 368 |
+
|
| 369 |
+
def list_available_topics() -> str:
|
| 370 |
+
"""
|
| 371 |
+
Lista los capítulos y temas indexados en la base de conocimiento.
|
| 372 |
+
|
| 373 |
+
Returns:
|
| 374 |
+
str: Lista organizada de capítulos por libro con sus secciones principales.
|
| 375 |
+
"""
|
| 376 |
+
output_parts = ["# 📚 Contenido Disponible en la Base de Conocimiento\n"]
|
| 377 |
+
|
| 378 |
+
for book_key, collection_name in [("ESL", "esl_chapters"), ("ISLP", "islp_chapters")]:
|
| 379 |
+
try:
|
| 380 |
+
collection = get_collection(collection_name)
|
| 381 |
+
all_data = collection.get(include=["metadatas"])
|
| 382 |
+
|
| 383 |
+
if not all_data["metadatas"]:
|
| 384 |
+
output_parts.append(f"\n## {book_key}: Sin datos\n")
|
| 385 |
+
continue
|
| 386 |
+
|
| 387 |
+
chapters = {}
|
| 388 |
+
for meta in all_data["metadatas"]:
|
| 389 |
+
chapter = meta.get("chapter", "Unknown")
|
| 390 |
+
section = meta.get("section", "")
|
| 391 |
+
if chapter not in chapters:
|
| 392 |
+
chapters[chapter] = set()
|
| 393 |
+
if section:
|
| 394 |
+
chapters[chapter].add(section)
|
| 395 |
+
|
| 396 |
+
output_parts.append(f"\n## 📗 {book_key}\n")
|
| 397 |
+
for chapter in sorted(chapters.keys()):
|
| 398 |
+
sections = sorted(chapters[chapter])
|
| 399 |
+
output_parts.append(f"\n### {chapter}\n")
|
| 400 |
+
if sections:
|
| 401 |
+
for sec in sections[:8]:
|
| 402 |
+
output_parts.append(f" - {sec}\n")
|
| 403 |
+
if len(sections) > 8:
|
| 404 |
+
output_parts.append(f" - ... y {len(sections) - 8} secciones más\n")
|
| 405 |
+
|
| 406 |
+
total = collection.count()
|
| 407 |
+
output_parts.append(f"\n**Total chunks indexados:** {total}\n")
|
| 408 |
+
|
| 409 |
+
except Exception as e:
|
| 410 |
+
output_parts.append(f"\n## {book_key}: ❌ Error ({e})\n")
|
| 411 |
+
|
| 412 |
+
return "\n".join(output_parts)
|