rag_utils_unsloth.py

import faiss
import pickle
import numpy as np
import torch
import os

from sentence_transformers import SentenceTransformer
from unsloth import FastLanguageModel

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():
    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)
    results = [documents[i] for i in indices[0]]
    return results

def generate_answer(question, context):
    model_id = "unsloth/mistral-7b-instruct-v0.1-bnb-4bit"

    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=model_id,
        max_seq_length=4096,
        dtype="float32",            # pour CPU uniquement
        load_in_4bit=True,
        device_map="auto"
    )

    tokenizer.pad_token = tokenizer.eos_token

    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)
    return tokenizer.decode(outputs[0], skip_special_tokens=True)