app.py

from fastapi import FastAPI, Request, Form
from fastapi.responses import HTMLResponse, JSONResponse
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates

import os
from tqdm import tqdm

# LangChain imports
from langchain_community.document_loaders import PyPDFLoader
from langchain_community.vectorstores import FAISS
from langchain_community.chat_models import ChatOpenAI
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.chains import RetrievalQA
from langchain_community.embeddings import HuggingFaceEmbeddings

# FastAPI setup
app = FastAPI()
app.mount("/static", StaticFiles(directory="."), name="static")
templates = Jinja2Templates(directory=".")

# Cargar e indexar los documentos PDF
def cargar_docs():
    all_docs = []
    for pdf_file in [
        "1 مساعد ممارس ملف المحور.pdf",
        "2 مساعد ممارس ملف المحور.pdf",
        "3 مساعد ممارس املف المحور.pdf",
        "4 مساعد ممارس ملف المحور.pdf",
        "ملف المحور 5 مساعد ممارس.pdf"
    ]:
        loader = PyPDFLoader(pdf_file)
        all_docs.extend(loader.load())
    return all_docs

docs = cargar_docs()

# Dividir en chunks
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1500, chunk_overlap=200)
split_docs = text_splitter.split_documents(docs)

# Debug: mostrar primeros chunks
for i, doc in enumerate(split_docs[:5]):
    print(f"Chunk {i+1}:\n{doc.page_content[:300]}\n{'-'*40}")

# ===================
# Embeddings árabes
# ===================
model_name = "asafaya/bert-base-arabic"
embeddings = HuggingFaceEmbeddings(model_name=model_name)

# Crear vectorstore
batch_size = 100
vectorstore = None

for i in tqdm(range(0, len(split_docs), batch_size)):
    batch = split_docs[i:i + batch_size]
    if vectorstore is None:
        vectorstore = FAISS.from_documents(batch, embeddings)
    else:
        vectorstore.add_documents(batch)

print(f"Se han indexado {len(split_docs)} chunks.")

# Cadena de respuesta
qa_chain = RetrievalQA.from_chain_type(
    llm=ChatOpenAI(temperature=0.2, model_name="gpt-4o-mini"),
    chain_type="stuff",
    retriever=vectorstore.as_retriever(search_kwargs={"k": 5})
)

@app.get("/", response_class=HTMLResponse)
async def read_root(request: Request):
    return templates.TemplateResponse("index.html", {"request": request})

@app.post("/preguntar")
async def preguntar(request: Request, pregunta: str = Form(...)):
    respuesta = qa_chain.run(pregunta)
    return JSONResponse({"respuesta": respuesta})

# Para modo local
import uvicorn

if __name__ == "__main__":
    uvicorn.run("app:app", host="0.0.0.0", port=7860)