Update rag.py
This commit is contained in:
parent
e26fc3e20d
commit
4f61b37db1
43
rag.py
43
rag.py
@ -16,27 +16,44 @@ import faiss, numpy as np, requests
|
||||
from FlagEmbedding import BGEM3FlagModel
|
||||
from rich import print
|
||||
|
||||
ROOT = Path("Corpus") # dossier racine des fiches (comme dans index.py)
|
||||
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:
|
||||
filepath = ROOT / m["path"]
|
||||
try:
|
||||
if filepath.exists() and filepath.suffix.lower() in {".md", ".markdown", ".txt"}:
|
||||
docs.append(filepath.read_text(encoding="utf-8"))
|
||||
else:
|
||||
docs.append(f"[passage manquant: {m['path']}]")
|
||||
except Exception as e:
|
||||
print(f"[dim]Erreur lecture {m['path']}: {e}[/]")
|
||||
docs.append(f"[erreur lecture: {m['path']}]")
|
||||
key = (m["file"], m["part"])
|
||||
docs.append(cache.get(key, "[passage manquant]"))
|
||||
|
||||
print(f"[dim]Passages rechargés : {len(docs)} (ordre conforme à l'index).[/]")
|
||||
|
||||
@ -63,7 +80,9 @@ try:
|
||||
|
||||
D, I = idx.search(q_emb.astype("float32").reshape(1, -1), K)
|
||||
hits = I[0]
|
||||
# contexte des passages trouvés
|
||||
# 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"
|
||||
@ -85,6 +104,6 @@ try:
|
||||
print("\n[dim]--- contexte utilisé ---[/]")
|
||||
for rank, idx_id in enumerate(hits, 1):
|
||||
m = meta[int(idx_id)]
|
||||
print(f"[{rank}] {m['path']} → {docs[int(idx_id)][:120]}…")
|
||||
print(f"[{rank}] {m['file']} · part {m['part']} → {docs[int(idx_id)][:120]}…")
|
||||
except Exception as e:
|
||||
print("[red]Erreur :", e)
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user