Spaces:
Running
Running
File size: 17,465 Bytes
80f110c c6a5264 80f110c f1ee128 80f110c c6a5264 80f110c c6a5264 80f110c c6a5264 8726e17 80f110c 8726e17 c6a5264 80f110c c6a5264 80f110c c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 80f110c c6a5264 80f110c f1ee128 c6a5264 80f110c 82a76f0 c6a5264 80f110c c6a5264 82a76f0 c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 80f110c 8726e17 c6a5264 8726e17 c6a5264 8726e17 80f110c c6a5264 80f110c c6a5264 b3f4ecb c6a5264 8726e17 a931c35 80f110c c6a5264 8726e17 80f110c c6a5264 8726e17 82a76f0 c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 82a76f0 c6a5264 82a76f0 c6a5264 80f110c c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 8726e17 c6a5264 80f110c c6a5264 80f110c c6a5264 80f110c c6a5264 80f110c c6a5264 903baeb c6a5264 903baeb c6a5264 8726e17 c6a5264 8726e17 c6a5264 b3f4ecb c6a5264 8726e17 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 |
# -*- coding: utf-8 -*-
from __future__ import annotations
import os
import io
import json
import time
import tarfile
import logging
import hashlib
from typing import Dict, Any, List, Tuple, Optional
import numpy as np
import faiss
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import JSONResponse, StreamingResponse
from pydantic import BaseModel
import gradio as gr
# =============================================================================
# LOGGING
# =============================================================================
LOG = logging.getLogger("remote-indexer-space")
if not LOG.handlers:
h = logging.StreamHandler()
h.setFormatter(logging.Formatter("%(asctime)s - %(levelname)s - %(message)s"))
LOG.addHandler(h)
LOG.setLevel(logging.INFO)
# =============================================================================
# CONFIG (via ENV)
# =============================================================================
PORT = int(os.getenv("PORT", "7860"))
DATA_ROOT = os.getenv("DATA_ROOT", "/tmp/data") # persistant dans le conteneur Space
os.makedirs(DATA_ROOT, exist_ok=True)
# Provider d'embeddings:
# - "dummy" : vecteurs aléatoires déterministes (très rapide)
# - "st" : Sentence-Transformers (CPU-friendly, simple)
# - "hf" : Transformers (AutoModel/AutoTokenizer, pooling manuel)
EMB_PROVIDER = os.getenv("EMB_PROVIDER", "dummy").strip().lower()
# Modèle embeddings (utilisé si provider != "dummy")
# Reco rapide et multilingue (FR ok) : paraphrase-multilingual-MiniLM-L12-v2 (dim=384)
EMB_MODEL = os.getenv("EMB_MODEL", "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2").strip()
# Batch d'encodage
EMB_BATCH = int(os.getenv("EMB_BATCH", "32"))
# Dimension par défaut (dummy) — pour st/hf on lit depuis le modèle
EMB_DIM = int(os.getenv("EMB_DIM", "128"))
# Cache global lazy
_ST_MODEL = None
_HF_TOKENIZER = None
_HF_MODEL = None
# =============================================================================
# JOB STATE
# =============================================================================
class JobState(BaseModel):
job_id: str
project_id: str
stage: str = "pending" # pending -> chunking -> embedding -> indexing -> done/failed
total_files: int = 0
total_chunks: int = 0
embedded: int = 0
indexed: int = 0
errors: List[str] = []
messages: List[str] = []
started_at: float = time.time()
finished_at: Optional[float] = None
JOBS: Dict[str, JobState] = {}
def _now() -> str:
return time.strftime("%H:%M:%S")
def _proj_dirs(project_id: str) -> Tuple[str, str, str]:
base = os.path.join(DATA_ROOT, project_id)
ds_dir = os.path.join(base, "dataset")
fx_dir = os.path.join(base, "faiss")
os.makedirs(ds_dir, exist_ok=True)
os.makedirs(fx_dir, exist_ok=True)
return base, ds_dir, fx_dir
def _add_msg(st: JobState, msg: str):
st.messages.append(f"[{_now()}] {msg}")
LOG.info("[%s] %s", st.job_id, msg)
def _set_stage(st: JobState, stage: str):
st.stage = stage
_add_msg(st, f"stage={stage}")
# =============================================================================
# UTILS
# =============================================================================
def _chunk_text(text: str, size: int = 200, overlap: int = 20) -> List[str]:
text = (text or "").replace("\r\n", "\n")
tokens = list(text)
if size <= 0:
return [text] if text else []
if overlap < 0:
overlap = 0
chunks = []
i = 0
while i < len(tokens):
j = min(i + size, len(tokens))
chunk = "".join(tokens[i:j]).strip()
if chunk:
chunks.append(chunk)
if j == len(tokens):
break
i = j - overlap if (j - overlap) > i else j
return chunks
def _l2_normalize(x: np.ndarray) -> np.ndarray:
n = np.linalg.norm(x, axis=1, keepdims=True) + 1e-12
return x / n
# ----------------------- PROVIDER: DUMMY --------------------------------------
def _emb_dummy(texts: List[str], dim: int = EMB_DIM) -> np.ndarray:
vecs = np.zeros((len(texts), dim), dtype="float32")
for i, t in enumerate(texts):
h = hashlib.sha1((t or "").encode("utf-8")).digest()
rng = np.random.default_rng(int.from_bytes(h[:8], "little", signed=False))
v = rng.standard_normal(dim).astype("float32")
vecs[i] = v / (np.linalg.norm(v) + 1e-9)
return vecs
# ----------------- PROVIDER: Sentence-Transformers ----------------------------
def _get_st_model():
global _ST_MODEL
if _ST_MODEL is None:
from sentence_transformers import SentenceTransformer
_ST_MODEL = SentenceTransformer(EMB_MODEL)
LOG.info(f"[st] modèle chargé: {EMB_MODEL}")
return _ST_MODEL
def _emb_st(texts: List[str]) -> np.ndarray:
model = _get_st_model()
vecs = model.encode(
texts,
batch_size=max(1, EMB_BATCH),
convert_to_numpy=True,
normalize_embeddings=True,
show_progress_bar=False,
).astype("float32")
return vecs
def _st_dim() -> int:
model = _get_st_model()
try:
return int(model.get_sentence_embedding_dimension())
except Exception:
# fallback : encode une phrase et lit la shape
v = model.encode(["dimension probe"], convert_to_numpy=True)
return int(v.shape[1])
# ----------------------- PROVIDER: Transformers (HF) --------------------------
def _get_hf_model():
global _HF_TOKENIZER, _HF_MODEL
if _HF_MODEL is None or _HF_TOKENIZER is None:
from transformers import AutoTokenizer, AutoModel
_HF_TOKENIZER = AutoTokenizer.from_pretrained(EMB_MODEL)
_HF_MODEL = AutoModel.from_pretrained(EMB_MODEL)
_HF_MODEL.eval()
LOG.info(f"[hf] modèle chargé: {EMB_MODEL}")
return _HF_TOKENIZER, _HF_MODEL
def _mean_pool(last_hidden_state: "np.ndarray", attention_mask: "np.ndarray") -> "np.ndarray":
# mean pooling masquée
mask = attention_mask[..., None].astype(last_hidden_state.dtype) # (b, t, 1)
summed = (last_hidden_state * mask).sum(axis=1) # (b, h)
counts = mask.sum(axis=1).clip(min=1e-9) # (b, 1)
return summed / counts
def _emb_hf(texts: List[str]) -> np.ndarray:
import torch
tok, mod = _get_hf_model()
all_vecs = []
bs = max(1, EMB_BATCH)
with torch.no_grad():
for i in range(0, len(texts), bs):
batch = texts[i:i+bs]
enc = tok(batch, padding=True, truncation=True, return_tensors="pt")
out = mod(**enc)
last = out.last_hidden_state # (b, t, h)
pooled = _mean_pool(last.numpy(), enc["attention_mask"].numpy()) # numpy
all_vecs.append(pooled.astype("float32"))
vecs = np.concatenate(all_vecs, axis=0)
return _l2_normalize(vecs)
def _hf_dim() -> int:
# essaie de lire hidden_size
try:
_, mod = _get_hf_model()
return int(getattr(mod.config, "hidden_size", 768))
except Exception:
return 768
# ---------------------------- DATASET / FAISS ---------------------------------
def _save_dataset(ds_dir: str, rows: List[Dict[str, Any]]):
os.makedirs(ds_dir, exist_ok=True)
data_path = os.path.join(ds_dir, "data.jsonl")
with open(data_path, "w", encoding="utf-8") as f:
for r in rows:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
meta = {"format": "jsonl", "columns": ["path", "text", "chunk_id"], "count": len(rows)}
with open(os.path.join(ds_dir, "meta.json"), "w", encoding="utf-8") as f:
json.dump(meta, f, ensure_ascii=False, indent=2)
def _load_dataset(ds_dir: str) -> List[Dict[str, Any]]:
data_path = os.path.join(ds_dir, "data.jsonl")
if not os.path.isfile(data_path):
return []
out = []
with open(data_path, "r", encoding="utf-8") as f:
for line in f:
try:
out.append(json.loads(line))
except Exception:
continue
return out
def _save_faiss(fx_dir: str, xb: np.ndarray, meta: Dict[str, Any]):
os.makedirs(fx_dir, exist_ok=True)
idx_path = os.path.join(fx_dir, "emb.faiss")
index = faiss.IndexFlatIP(xb.shape[1]) # cosine ~ inner product si normalisé
index.add(xb)
faiss.write_index(index, idx_path)
with open(os.path.join(fx_dir, "meta.json"), "w", encoding="utf-8") as f:
json.dump(meta, f, ensure_ascii=False, indent=2)
def _load_faiss(fx_dir: str) -> faiss.Index:
idx_path = os.path.join(fx_dir, "emb.faiss")
if not os.path.isfile(idx_path):
raise FileNotFoundError(f"FAISS index introuvable: {idx_path}")
return faiss.read_index(idx_path)
def _tar_dir_to_bytes(dir_path: str) -> bytes:
bio = io.BytesIO()
with tarfile.open(fileobj=bio, mode="w:gz") as tar:
tar.add(dir_path, arcname=os.path.basename(dir_path))
bio.seek(0)
return bio.read()
# =============================================================================
# FASTAPI
# =============================================================================
fastapi_app = FastAPI(title="remote-indexer", version="2.0.0")
fastapi_app.add_middleware(
CORSMiddleware,
allow_origins=["*"], allow_credentials=True, allow_methods=["*"], allow_headers=["*"],
)
class FileItem(BaseModel):
path: str
text: str
class IndexRequest(BaseModel):
project_id: str
files: List[FileItem]
chunk_size: int = 200
overlap: int = 20
batch_size: int = 32
store_text: bool = True
@fastapi_app.get("/health")
def health():
info = {
"ok": True,
"service": "remote-indexer",
"provider": EMB_PROVIDER,
"model": EMB_MODEL if EMB_PROVIDER != "dummy" else None
}
return info
@fastapi_app.get("/")
def root_redirect():
return {"ok": True, "service": "remote-indexer", "ui": "/ui"}
@fastapi_app.post("/index")
def index(req: IndexRequest):
job_id = hashlib.sha1(f"{req.project_id}{time.time()}".encode()).hexdigest()[:12]
st = JobState(job_id=job_id, project_id=req.project_id, stage="pending", messages=[])
JOBS[job_id] = st
_add_msg(st, f"Job {job_id} créé pour project {req.project_id}")
_add_msg(st, f"Index start project={req.project_id} files={len(req.files)} chunk_size={req.chunk_size} overlap={req.overlap} batch_size={req.batch_size} store_text={req.store_text} provider={EMB_PROVIDER} model={EMB_MODEL if EMB_PROVIDER!='dummy' else '-'}")
try:
base, ds_dir, fx_dir = _proj_dirs(req.project_id)
# 1) Chunking
_set_stage(st, "chunking")
rows: List[Dict[str, Any]] = []
st.total_files = len(req.files)
for it in req.files:
txt = it.text or ""
chunks = _chunk_text(txt, size=req.chunk_size, overlap=req.overlap)
_add_msg(st, f"{it.path}: len(text)={len(txt)} chunks={len(chunks)}")
for ci, ck in enumerate(chunks):
rows.append({"path": it.path, "text": ck, "chunk_id": ci})
st.total_chunks = len(rows)
_add_msg(st, f"Total chunks = {st.total_chunks}")
# 2) Embedding
_set_stage(st, "embedding")
if EMB_PROVIDER == "dummy":
xb = _emb_dummy([r["text"] for r in rows], dim=EMB_DIM)
dim = xb.shape[1]
elif EMB_PROVIDER == "st":
xb = _emb_st([r["text"] for r in rows])
dim = xb.shape[1]
else: # "hf"
xb = _emb_hf([r["text"] for r in rows])
dim = xb.shape[1]
st.embedded = xb.shape[0]
_add_msg(st, f"Embeddings {st.embedded}/{st.total_chunks}")
_add_msg(st, f"Embeddings dim={dim}")
# 3) Sauvegarde dataset (texte)
_save_dataset(ds_dir, rows)
_add_msg(st, f"Dataset (sans index) sauvegardé dans {ds_dir}")
# 4) FAISS
_set_stage(st, "indexing")
faiss_meta = {
"dim": int(dim),
"count": int(xb.shape[0]),
"provider": EMB_PROVIDER,
"model": EMB_MODEL if EMB_PROVIDER != "dummy" else None
}
_save_faiss(fx_dir, xb, meta=faiss_meta)
st.indexed = int(xb.shape[0])
_add_msg(st, f"FAISS écrit sur {os.path.join(fx_dir, 'emb.faiss')}")
_add_msg(st, f"OK — dataset+index prêts (projet={req.project_id})")
_set_stage(st, "done")
st.finished_at = time.time()
return {"job_id": job_id}
except Exception as e:
LOG.exception("index failed")
st.errors.append(str(e))
_add_msg(st, f"❌ Exception: {e}")
st.stage = "failed"
st.finished_at = time.time()
raise HTTPException(status_code=500, detail=str(e))
@fastapi_app.get("/status/{job_id}")
def status(job_id: str):
st = JOBS.get(job_id)
if not st:
raise HTTPException(status_code=404, detail="job inconnu")
return JSONResponse(st.model_dump())
class SearchRequest(BaseModel):
project_id: str
query: str
k: int = 5
@fastapi_app.post("/search")
def search(req: SearchRequest):
base, ds_dir, fx_dir = _proj_dirs(req.project_id)
rows = _load_dataset(ds_dir)
if not rows:
raise HTTPException(status_code=404, detail="dataset introuvable (index pas encore construit ?)")
# Embedding de la requête avec le MÊME provider
if EMB_PROVIDER == "dummy":
q = _emb_dummy([req.query], dim=EMB_DIM)[0:1, :]
elif EMB_PROVIDER == "st":
q = _emb_st([req.query])[0:1, :]
else:
q = _emb_hf([req.query])[0:1, :]
# FAISS
index = _load_faiss(fx_dir)
if index.d != q.shape[1]:
raise HTTPException(status_code=500, detail=f"dim incompatibles: index.d={index.d} vs query={q.shape[1]}")
scores, ids = index.search(q, int(max(1, req.k)))
ids = ids[0].tolist()
scores = scores[0].tolist()
out = []
for idx, sc in zip(ids, scores):
if idx < 0 or idx >= len(rows):
continue
r = rows[idx]
out.append({"path": r.get("path"), "text": r.get("text"), "score": float(sc)})
return {"results": out}
# ----------- ARTIFACTS EXPORT -----------
@fastapi_app.get("/artifacts/{project_id}/dataset")
def download_dataset(project_id: str):
base, ds_dir, _ = _proj_dirs(project_id)
if not os.path.isdir(ds_dir):
raise HTTPException(status_code=404, detail="Dataset introuvable")
buf = _tar_dir_to_bytes(ds_dir)
headers = {"Content-Disposition": f'attachment; filename="{project_id}_dataset.tgz"'}
return StreamingResponse(io.BytesIO(buf), media_type="application/gzip", headers=headers)
@fastapi_app.get("/artifacts/{project_id}/faiss")
def download_faiss(project_id: str):
base, _, fx_dir = _proj_dirs(project_id)
if not os.path.isdir(fx_dir):
raise HTTPException(status_code=404, detail="FAISS introuvable")
buf = _tar_dir_to_bytes(fx_dir)
headers = {"Content-Disposition": f'attachment; filename="{project_id}_faiss.tgz"'}
return StreamingResponse(io.BytesIO(buf), media_type="application/gzip", headers=headers)
# =============================================================================
# GRADIO UI (facultatif)
# =============================================================================
def _ui_index(project_id: str, sample_text: str):
files = [{"path": "sample.txt", "text": sample_text}]
from pydantic import ValidationError
try:
req = IndexRequest(project_id=project_id, files=[FileItem(**f) for f in files])
except ValidationError as e:
return f"Erreur: {e}"
try:
res = index(req)
return f"Job lancé: {res['job_id']}"
except Exception as e:
return f"Erreur index: {e}"
def _ui_search(project_id: str, query: str, k: int):
try:
res = search(SearchRequest(project_id=project_id, query=query, k=int(k)))
return json.dumps(res, ensure_ascii=False, indent=2)
except Exception as e:
return f"Erreur search: {e}"
with gr.Blocks(title="Remote Indexer (FAISS)", analytics_enabled=False) as ui:
gr.Markdown("## Remote Indexer — demo UI (API: `/index`, `/status/{job}`, `/search`, `/artifacts/...`).")
gr.Markdown(f"**Provider**: `{EMB_PROVIDER}` — **Model**: `{EMB_MODEL if EMB_PROVIDER!='dummy' else '-'}'")
with gr.Tab("Index"):
pid = gr.Textbox(label="Project ID", value="DEEPWEB")
sample = gr.Textbox(label="Texte d’exemple", value="Alpha bravo charlie delta echo foxtrot.", lines=4)
btn = gr.Button("Lancer index (sample)")
out = gr.Textbox(label="Résultat")
btn.click(_ui_index, inputs=[pid, sample], outputs=[out])
with gr.Tab("Search"):
pid2 = gr.Textbox(label="Project ID", value="DEEPWEB")
q = gr.Textbox(label="Query", value="alpha")
k = gr.Slider(1, 20, value=5, step=1, label="k")
btn2 = gr.Button("Rechercher")
out2 = gr.Code(label="Résultats")
btn2.click(_ui_search, inputs=[pid2, q, k], outputs=[out2])
fastapi_app = gr.mount_gradio_app(fastapi_app, ui, path="/ui")
# =============================================================================
# MAIN
# =============================================================================
if __name__ == "__main__":
import uvicorn
LOG.info("Démarrage Uvicorn sur 0.0.0.0:%s (UI_PATH=/ui)", PORT)
uvicorn.run(fastapi_app, host="0.0.0.0", port=PORT) |