rag_utils.py

import os
import faiss
import pickle
import numpy as np
import torch
from sentence_transformers import SentenceTransformer
from transformers import AutoModelForCausalLM, AutoTokenizer

def load_faiss_index(index_path="faiss_index/faiss_index.faiss", doc_path="faiss_index/documents.pkl"):
    index = faiss.read_index(index_path)
    with open(doc_path, "rb") as f:
        documents = pickle.load(f)
    return index, documents

def get_embedding_model():
    # Pas besoin de token ici, modèle public
    print("✅ Chargement de l'encodeur multi-qa-MiniLM-L6-cos-v1")
    return SentenceTransformer("sentence-transformers/multi-qa-MiniLM-L6-cos-v1")

def query_index(question, index, documents, model, k=3):
    question_embedding = model.encode([question])
    _, indices = index.search(np.array(question_embedding).astype("float32"), k)
    return [documents[i] for i in indices[0]]

def generate_answer(question, context):
    token = os.getenv("HUGGINGFACE")  # requis pour Mistral
    model_id = "mgoogle/flan-t5-base"

    tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
    tokenizer.pad_token = tokenizer.eos_token

    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        token=token,
        device_map="auto",
        torch_dtype=torch.float16
    )

    prompt = f"Voici un contexte :\n{context}\n\nQuestion : {question}\nRéponse :"
    inputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True).to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=256, pad_token_id=tokenizer.eos_token_id)
    print("🔍 Contexte utilisé pour la génération :")
    print(context[:500])
    return tokenizer.decode(outputs[0], skip_special_tokens=True)