Stéphan
a3608353a2
Enhance split function to detect and preserve Markdown tables when chunking text. Tables are now kept intact by forcing splits before and after table content. Also increase K value from 10 to 30 in rag.py to provide more passages to the LLM.
110 lines
4.0 KiB
Python
110 lines
4.0 KiB
Python
#!/usr/bin/env python3
|
||
"""
|
||
RAG interactif – version alignée sur l'index
|
||
-------------------------------------------
|
||
• Utilise corpus.idx + corpus.meta.json pour connaître l'ordre exact des passages.
|
||
• Recharge **uniquement** les textes correspondants en gardant cet ordre – ainsi, plus
|
||
d'erreur d'index out‑of‑range quelle que soit la découpe.
|
||
• Recherche FAISS (top‑k=4) + génération via mistral7b-fast (Ollama).
|
||
"""
|
||
|
||
import json, readline, re
|
||
from pathlib import Path
|
||
from collections import defaultdict
|
||
|
||
import faiss, numpy as np, requests
|
||
from FlagEmbedding import BGEM3FlagModel
|
||
from rich import print
|
||
|
||
ROOT = Path("Fiches") # dossier racine des fiches
|
||
K = 30 # nombre de passages remis au LLM
|
||
|
||
# ------------------ utilitaires de découpe identiques à l'index -------------
|
||
CHUNK, OVERLAP = 800, 100 # garder cohérent avec index.py
|
||
|
||
def split(text: str):
|
||
sents = re.split(r"(?<=[.!?]) +", text)
|
||
buf, out = [], []
|
||
for s in sents:
|
||
buf.append(s)
|
||
if len(" ".join(buf).split()) > CHUNK:
|
||
out.append(" ".join(buf))
|
||
buf = buf[-OVERLAP:]
|
||
if buf:
|
||
out.append(" ".join(buf))
|
||
return out
|
||
|
||
# ------------------- charger meta et reconstruire passages ------------------
|
||
meta_path = Path("corpus.meta.json")
|
||
if not meta_path.exists():
|
||
raise SystemExit("corpus.meta.json introuvable – lancez d'abord index.py")
|
||
meta = json.load(meta_path.open())
|
||
|
||
# mapping (file, part) -> chunk text
|
||
cache: dict[tuple[str, int], str] = {}
|
||
for fp in sorted(ROOT.rglob("*")):
|
||
if fp.suffix.lower() not in {".md", ".markdown", ".txt"}:
|
||
continue
|
||
chunks = split(fp.read_text(encoding="utf-8"))
|
||
for i, ch in enumerate(chunks):
|
||
cache[(fp.name, i)] = ch
|
||
|
||
# reconstruire docs dans le même ordre que l'index ---------------------------
|
||
docs = []
|
||
for m in meta:
|
||
key = (m["file"], m["part"])
|
||
docs.append(cache.get(key, "[passage manquant]"))
|
||
|
||
print(f"[dim]Passages rechargés : {len(docs)} (ordre conforme à l'index).[/]")
|
||
|
||
# ---------------- FAISS + modèle embeddings --------------------------------
|
||
idx = faiss.read_index("corpus.idx")
|
||
model = BGEM3FlagModel("BAAI/bge-m3", device="cpu")
|
||
|
||
# ---------------- boucle interactive ---------------------------------------
|
||
print("RAG prêt. Posez vos questions ! (Ctrl‑D pour sortir)")
|
||
try:
|
||
while True:
|
||
try:
|
||
q = input("❓ > ").strip()
|
||
if not q:
|
||
continue
|
||
except (EOFError, KeyboardInterrupt):
|
||
print("\nBye."); break
|
||
|
||
emb = model.encode([q])
|
||
if isinstance(emb, dict):
|
||
# récupère le 1er ndarray trouvé
|
||
emb = next(v for v in emb.values() if isinstance(v, np.ndarray))
|
||
q_emb = emb[0] / np.linalg.norm(emb[0])
|
||
|
||
D, I = idx.search(q_emb.astype("float32").reshape(1, -1), K)
|
||
hits = I[0]
|
||
# réordonne pour mettre en tête les passages contenant “Seuil”
|
||
hits = sorted(hits, key=lambda i: "Seuil" not in docs[int(i)])
|
||
|
||
context = "\n\n".join(docs[int(i)] for i in hits[:K])
|
||
prompt = (
|
||
"<system>Réponds en français, de façon précise, et uniquement à partir du contexte fourni. Si l'information n'est pas dans le contexte, réponds : 'Je ne sais pas'.</system>\n"
|
||
f"<context>{context}</context>\n<user>{q}</user>"
|
||
)
|
||
|
||
def ask_llm(p):
|
||
r = requests.post("http://127.0.0.1:11434/api/generate", json={
|
||
"model": "mistral7b-fast",
|
||
"prompt": p,
|
||
"stream": False,
|
||
"options": {"temperature": 0.0, "num_predict": 512}
|
||
}, timeout=300)
|
||
return r.json()["response"]
|
||
|
||
print("\n[bold]Réponse :[/]")
|
||
print(ask_llm(prompt))
|
||
|
||
print("\n[dim]--- contexte utilisé ---[/]")
|
||
for rank, idx_id in enumerate(hits, 1):
|
||
m = meta[int(idx_id)]
|
||
print(f"[{rank}] {m['file']} · part {m['part']} → {docs[int(idx_id)][:120]}…")
|
||
except Exception as e:
|
||
print("[red]Erreur :", e)
|