app.py

import os, urllib.request
from contextlib import asynccontextmanager
from dotenv import load_dotenv
from fastapi import FastAPI
from pydantic import BaseModel
from langchain_community.document_loaders import PyPDFLoader
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_nvidia_ai_endpoints import ChatNVIDIA, NVIDIAEmbeddings
from langchain_chroma import Chroma
from langchain_core.prompts import ChatPromptTemplate

load_dotenv()

DOCS_DIR = "/tmp/documents/"

rag_state = {}


def telecharger_documents():
    os.makedirs(DOCS_DIR, exist_ok=True)
    dest = os.path.join(DOCS_DIR, "CONVENTION_SYNTEC.pdf")
    if os.path.exists(dest):
        print(f"Le fichier {dest} existe deja — telechargement ignore.")
    else:
        print("Telechargement de CONVENTION_SYNTEC.pdf...")
        urllib.request.urlretrieve(
            "https://github.com/archiducarmel/SupDeVinci_M1_MachineLearning_DeepLearning/releases/download/datas/CONVENTION_SYNTEC.pdf",
            dest,
        )
        print("OK.")
    size_kb = os.path.getsize(dest) / 1024
    print(f"\n✅ {dest} ({size_kb:.0f} Ko) pret dans ./documents/")


def extract_answer(response):
    text = (response.content or "").strip()
    if not text:
        text = (response.additional_kwargs.get("reasoning_content", "") or "").strip()
    return text


def format_docs(docs):
    return "\n\n".join(
        f"[{d.metadata.get('source', '?').split('/')[-1]} — page {d.metadata.get('page')}] {d.page_content}"
        for d in docs
    )


@asynccontextmanager
async def lifespan(app: FastAPI):
    telecharger_documents()

    all_docs = []
    for f in sorted(os.listdir(DOCS_DIR)):
        if f.endswith(".pdf"):
            all_docs.extend(PyPDFLoader(os.path.join(DOCS_DIR, f)).load())
    chunks = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100).split_documents(all_docs)

    embeddings = NVIDIAEmbeddings(model="nvidia/llama-nemotron-embed-1b-v2", truncate="NONE")
    vector_store = Chroma.from_documents(chunks, embeddings)

    prompt = ChatPromptTemplate.from_template(
        "Tu es un assistant RH expert de la convention collective Syntec. "
        "Réponds à la QUESTION en t'appuyant UNIQUEMENT sur le CONTEXTE ci-dessous.\n"
        "Si l'information n'y figure pas, réponds exactement : « Je ne sais pas ».\n"
        "Sois concis et cite la source (document et numéro de page).\n\n"
        "CONTEXTE :\n{context}\n\nQUESTION : {input}"
    )
    llm = ChatNVIDIA(model="openai/gpt-oss-120b", temperature=0.2, max_completion_tokens=2048)

    rag_state["retriever"] = vector_store.as_retriever(search_kwargs={"k": 3})
    rag_state["generation"] = prompt | llm | extract_answer

    yield

    rag_state.clear()


app = FastAPI(title="Assistant RH — Convention Syntec — API RAG", lifespan=lifespan)


class QuestionIn(BaseModel):
    question: str


def rag_answer(question):
    docs = rag_state["retriever"].invoke(question)
    answer = rag_state["generation"].invoke({"context": format_docs(docs), "input": question})
    return {"answer": answer, "context": docs}


@app.get("/")
def health():
    return {"status": "ok"}


@app.post("/ask")
def ask(payload: QuestionIn):
    result = rag_answer(payload.question)
    sources = [
        {"document": d.metadata.get("source", "?").split("/")[-1], "page": d.metadata.get("page")}
        for d in result["context"]
    ]
    return {"answer": result["answer"], "sources": sources}