rag_utils.py

import faiss
import pickle
import numpy as np
import re
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download
from llama_cpp import Llama

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)
    return [documents[i] for i in indices[0]]

def nettoyer_context(context):
    context = re.sub(r"\[\'(.*?)\'\]", r"\1", context)
    context = context.replace("None", "")
    return context

def generate_answer(question, context):
    model_file = hf_hub_download(
        repo_id="TheBloke/Mistral-7B-Instruct-v0.1-GGUF",
        filename="mistral-7b-instruct-v0.1.Q4_K_M.gguf"
    )

    llm = Llama(
        model_path=model_file,
        n_ctx=2048,
        n_threads=6,
        verbose=False
    )

    prompt = f"""Voici des informations sur des établissements et formations :

{context}

Formule ta réponse comme un conseiller d’orientation bienveillant, de manière fluide et naturelle, sans énumérations brutes.

Question : {question}
Réponse :
"""

    output = llm(prompt, max_tokens=128, stop=["</s>"])
    return output["choices"][0]["text"].strip()