DeepIndex / main.py
chouchouvs's picture
Update main.py
40bbd1a verified
raw
history blame
23 kB
# -*- coding: utf-8 -*-
from __future__ import annotations
import os, time, uuid, logging, random
from typing import List, Optional, Dict, Any, Tuple
import numpy as np
import requests
from fastapi import FastAPI, BackgroundTasks, Header, HTTPException, Query
from pydantic import BaseModel, Field
# Qdrant (optionnel si VECTOR_STORE=memory)
try:
from qdrant_client import QdrantClient
from qdrant_client.http.models import VectorParams, Distance, PointStruct
except Exception: # si non installé, on retombe en mémoire
QdrantClient = None
VectorParams = Distance = PointStruct = None
# ---------- logging ----------
logging.basicConfig(level=logging.INFO, format="%(levelname)s:%(name)s:%(message)s")
LOG = logging.getLogger("remote_indexer")
# ---------- ENV (config) ----------
# Choix du store: "qdrant" (par défaut) ou "memory"
VECTOR_STORE = os.getenv("VECTOR_STORE", "qdrant").strip().lower()
# Ordre des backends d'embeddings à essayer. Par défaut: DeepInfra, puis HF.
DEFAULT_BACKENDS = "deepinfra,hf"
EMB_BACKEND_ORDER = [s.strip().lower()
for s in os.getenv("EMB_BACKEND_ORDER", os.getenv("EMB_BACKEND", DEFAULT_BACKENDS)).split(",")
if s.strip()]
ALLOW_DI_AUTOFALLBACK = os.getenv("ALLOW_DI_AUTOFALLBACK", "true").lower() in ("1","true","yes","on")
# HF Inference API
HF_TOKEN = os.getenv("HF_API_TOKEN", "").strip()
HF_MODEL = os.getenv("HF_EMBED_MODEL", "sentence-transformers/all-MiniLM-L6-v2").strip()
HF_API_URL_USER = os.getenv("HF_API_URL", "").strip()
HF_API_URL_PIPELINE = os.getenv("HF_API_URL_PIPELINE", "").strip()
HF_API_URL_MODELS = os.getenv("HF_API_URL_MODELS", "").strip()
if HF_API_URL_USER:
if "/pipeline" in HF_API_URL_USER:
HF_API_URL_PIPELINE = HF_API_URL_USER
else:
HF_API_URL_MODELS = HF_API_URL_USER
HF_URL_PIPELINE = (HF_API_URL_PIPELINE or f"https://api-inference.huggingface.co/pipeline/feature-extraction/{HF_MODEL}")
HF_URL_MODELS = (HF_API_URL_MODELS or f"https://api-inference.huggingface.co/models/{HF_MODEL}")
HF_TIMEOUT = float(os.getenv("EMB_TIMEOUT_SEC", "120"))
HF_WAIT = os.getenv("HF_WAIT_FOR_MODEL", "true").lower() in ("1","true","yes","on")
HF_PIPELINE_FIRST = os.getenv("HF_PIPELINE_FIRST", "true").lower() in ("1","true","yes","on")
# DeepInfra (OpenAI-compatible embeddings)
DI_TOKEN = os.getenv("DEEPINFRA_API_KEY", "").strip()
DI_MODEL = os.getenv("DEEPINFRA_EMBED_MODEL", "BAAI/bge-m3").strip()
DI_URL = os.getenv("DEEPINFRA_EMBED_URL", "https://api.deepinfra.com/v1/openai/embeddings").strip()
DI_TIMEOUT = float(os.getenv("EMB_TIMEOUT_SEC", "120"))
# Retries embeddings
RETRY_MAX = int(os.getenv("EMB_RETRY_MAX", "6"))
RETRY_BASE_SEC = float(os.getenv("EMB_RETRY_BASE", "1.5"))
RETRY_JITTER = float(os.getenv("EMB_RETRY_JITTER", "0.35"))
# Qdrant
QDRANT_URL = os.getenv("QDRANT_URL", "http://localhost:6333").strip()
QDRANT_API = os.getenv("QDRANT_API_KEY", "").strip()
# Auth d’API du service (simple header)
AUTH_TOKEN = os.getenv("REMOTE_INDEX_TOKEN", "").strip()
LOG.info(f"Embeddings backend order = {EMB_BACKEND_ORDER}")
LOG.info(f"HF pipeline URL = {HF_URL_PIPELINE}")
LOG.info(f"HF models URL = {HF_URL_MODELS}")
LOG.info(f"VECTOR_STORE = {VECTOR_STORE}")
if "hf" in EMB_BACKEND_ORDER and not HF_TOKEN:
LOG.warning("HF_API_TOKEN manquant — tentatives HF échoueront.")
if "deepinfra" in EMB_BACKEND_ORDER and not DI_TOKEN:
LOG.warning("DEEPINFRA_API_KEY manquant — tentatives DeepInfra échoueront.")
# ---------- Vector store abstraction ----------
class VectorStoreBase:
def ensure_collection(self, name: str, dim: int): ...
def upsert(self, name: str, vectors: np.ndarray, payloads: List[dict]) -> int: ...
def search(self, name: str, query_vec: np.ndarray, limit: int):
"""return list of objects with .score and .payload"""
...
def wipe(self, name: str): ...
class MemoryHit:
def __init__(self, score: float, payload: dict):
self.score = score
self.payload = payload
class MemoryStore(VectorStoreBase):
"""Simple store en mémoire (cosine sur vecteurs normalisés). Persistance: vie du process."""
def __init__(self):
self.data: Dict[str, Dict[str, Any]] = {} # {col: {"dim": d, "vecs": np.ndarray [N,d], "payloads": List[dict]}}
LOG.warning("Vector store: MEMORY (fallback). Les données sont volatiles (perdues au restart).")
def ensure_collection(self, name: str, dim: int):
col = self.data.get(name)
if not col:
self.data[name] = {"dim": dim, "vecs": np.zeros((0, dim), dtype=np.float32), "payloads": []}
def upsert(self, name: str, vectors: np.ndarray, payloads: List[dict]) -> int:
self.ensure_collection(name, vectors.shape[1])
col = self.data[name]
if vectors.ndim != 2 or vectors.shape[1] != col["dim"]:
raise RuntimeError(f"MemoryStore: bad shape {vectors.shape}, expected (*,{col['dim']})")
col["vecs"] = np.vstack([col["vecs"], vectors.astype(np.float32)])
col["payloads"].extend(payloads)
return vectors.shape[0]
def search(self, name: str, query_vec: np.ndarray, limit: int):
col = self.data.get(name)
if not col or col["vecs"].shape[0] == 0:
return []
V = col["vecs"] # [N,d], déjà normalisés
q = query_vec.reshape(1, -1) # [1,d]
scores = (V @ q.T).ravel() # cos sim
idx = np.argsort(-scores)[:limit]
return [MemoryHit(float(scores[i]), col["payloads"][i]) for i in idx]
def wipe(self, name: str):
if name in self.data:
del self.data[name]
class QdrantStore(VectorStoreBase):
def __init__(self, url: str, api_key: Optional[str]):
if QdrantClient is None:
raise RuntimeError("qdrant-client non installé.")
self.client = QdrantClient(url=url, api_key=api_key if api_key else None)
# ping rapide
try:
_ = self.client.get_collections()
LOG.info("Connecté à Qdrant.")
except Exception as e:
raise RuntimeError(f"Connexion Qdrant impossible: {e}")
def ensure_collection(self, name: str, dim: int):
try:
self.client.get_collection(name); return
except Exception:
pass
self.client.create_collection(
collection_name=name,
vectors_config=VectorParams(size=dim, distance=Distance.COSINE),
)
def upsert(self, name: str, vectors: np.ndarray, payloads: List[dict]) -> int:
points = [
PointStruct(id=None, vector=v.tolist(), payload=payloads[i])
for i, v in enumerate(vectors)
]
self.client.upsert(collection_name=name, points=points)
return len(points)
def search(self, name: str, query_vec: np.ndarray, limit: int):
res = self.client.search(collection_name=name, query_vector=query_vec.tolist(), limit=limit)
return res
def wipe(self, name: str):
self.client.delete_collection(name)
# Sélection / auto-fallback du store
STORE: VectorStoreBase
def _init_store() -> VectorStoreBase:
prefer = VECTOR_STORE
if prefer == "memory":
return MemoryStore()
# prefer qdrant
try:
return QdrantStore(QDRANT_URL, QDRANT_API if QDRANT_API else None)
except Exception as e:
LOG.error(f"Qdrant indisponible ({e}) — fallback en mémoire.")
return MemoryStore()
STORE = _init_store()
# ---------- Pydantic ----------
class FileIn(BaseModel):
path: str
text: str
class IndexRequest(BaseModel):
project_id: str = Field(..., min_length=1)
files: List[FileIn]
chunk_size: int = 1200
overlap: int = 200
batch_size: int = 8
store_text: bool = True
class QueryRequest(BaseModel):
project_id: str
query: str
top_k: int = 6
# ---------- Jobs store ----------
JOBS: Dict[str, Dict[str, Any]] = {}
def _append_log(job_id: str, line: str):
job = JOBS.get(job_id)
if job: job["logs"].append(line)
def _set_status(job_id: str, status: str):
job = JOBS.get(job_id)
if job: job["status"] = status
def _auth(x_auth: Optional[str]):
if AUTH_TOKEN and (x_auth or "") != AUTH_TOKEN:
raise HTTPException(status_code=401, detail="Unauthorized")
# ---------- Helpers retry ----------
def _retry_sleep(attempt: int):
back = (RETRY_BASE_SEC ** attempt)
jitter = 1.0 + random.uniform(-RETRY_JITTER, RETRY_JITTER)
return max(0.25, back * jitter)
def _with_task_param(url: str, task: str = "feature-extraction") -> str:
return url + ("&" if "?" in url else "?") + f"task={task}"
# ---------- HF embeddings ----------
def _hf_http(url: str, payload: Dict[str, Any], headers_extra: Optional[Dict[str, str]] = None) -> Tuple[np.ndarray, int]:
if not HF_TOKEN:
raise RuntimeError("HF_API_TOKEN manquant (backend=hf).")
headers = {
"Authorization": f"Bearer {HF_TOKEN}",
"Content-Type": "application/json",
"Accept": "application/json",
}
if HF_WAIT:
payload.setdefault("options", {})["wait_for_model"] = True
headers["X-Wait-For-Model"] = "true"
headers["X-Use-Cache"] = "true"
if headers_extra:
headers.update(headers_extra)
r = requests.post(url, headers=headers, json=payload, timeout=HF_TIMEOUT)
size = int(r.headers.get("Content-Length", "0"))
if r.status_code >= 400:
LOG.error(f"HF error {r.status_code}: {r.text[:1000]}")
r.raise_for_status()
data = r.json()
arr = np.array(data, dtype=np.float32)
if arr.ndim == 3:
arr = arr.mean(axis=1)
elif arr.ndim == 1:
arr = arr.reshape(1, -1)
if arr.ndim != 2:
raise RuntimeError(f"HF: unexpected embeddings shape: {arr.shape}")
norms = np.linalg.norm(arr, axis=1, keepdims=True) + 1e-12
arr = arr / norms
return arr.astype(np.float32), size
def _hf_post_embeddings_once(batch: List[str]) -> Tuple[np.ndarray, int]:
payload: Dict[str, Any] = {"inputs": (batch if len(batch) > 1 else batch[0])}
urls = [HF_URL_PIPELINE, HF_URL_MODELS] if HF_PIPELINE_FIRST else [HF_URL_MODELS, HF_URL_PIPELINE]
last_exc: Optional[Exception] = None
for idx, url in enumerate(urls, 1):
try:
if "/models/" in url:
return _hf_http(url, payload, headers_extra={"X-Task": "feature-extraction"})
else:
return _hf_http(url, payload, headers_extra=None)
except requests.HTTPError as he:
code = he.response.status_code if he.response is not None else 0
body = he.response.text if he.response is not None else ""
last_exc = he
if code in (404, 405, 501) and idx < len(urls):
LOG.warning(f"HF endpoint {url} non dispo ({code}), fallback vers alternative ...")
continue
if "/models/" in url and "SentenceSimilarityPipeline" in (body or ""):
try:
forced_url = _with_task_param(url, "feature-extraction")
LOG.warning("HF MODELS a choisi Similarity -> retry avec %s + X-Task", forced_url)
return _hf_http(forced_url, payload, headers_extra={"X-Task": "feature-extraction"})
except Exception as he2:
last_exc = he2
raise
except Exception as e:
last_exc = e
raise
raise RuntimeError(f"HF: aucun endpoint utilisable ({last_exc})")
# ---------- DeepInfra embeddings ----------
def _di_post_embeddings_once(batch: List[str]) -> Tuple[np.ndarray, int]:
if not DI_TOKEN:
raise RuntimeError("DEEPINFRA_API_KEY manquant (backend=deepinfra).")
headers = {"Authorization": f"Bearer {DI_TOKEN}", "Content-Type": "application/json", "Accept": "application/json"}
payload = {"model": DI_MODEL, "input": batch}
r = requests.post(DI_URL, headers=headers, json=payload, timeout=DI_TIMEOUT)
size = int(r.headers.get("Content-Length", "0"))
if r.status_code >= 400:
LOG.error(f"DeepInfra error {r.status_code}: {r.text[:1000]}")
r.raise_for_status()
js = r.json()
data = js.get("data")
if not isinstance(data, list) or not data:
raise RuntimeError(f"DeepInfra embeddings: réponse invalide {js}")
embs = [d.get("embedding") for d in data]
arr = np.asarray(embs, dtype=np.float32)
if arr.ndim != 2:
raise RuntimeError(f"DeepInfra: unexpected embeddings shape: {arr.shape}")
norms = np.linalg.norm(arr, axis=1, keepdims=True) + 1e-12
arr = arr / norms
return arr.astype(np.float32), size
# ---------- Retry orchestrator ----------
def _retry_sleep(attempt: int):
back = (RETRY_BASE_SEC ** attempt)
jitter = 1.0 + random.uniform(-RETRY_JITTER, RETRY_JITTER)
return max(0.25, back * jitter)
def _call_with_retries(func, batch: List[str], label: str, job_id: Optional[str] = None) -> Tuple[np.ndarray, int]:
last_exc = None
for attempt in range(RETRY_MAX):
try:
if job_id:
_append_log(job_id, f"{label}: try {attempt+1}/{RETRY_MAX} (batch={len(batch)})")
return func(batch)
except requests.HTTPError as he:
code = he.response.status_code if he.response is not None else "HTTP"
retriable = code in (429, 500, 502, 503, 504)
if not retriable:
raise
sleep_s = _retry_sleep(attempt)
msg = f"{label}: HTTP {code}, retry in {sleep_s:.1f}s"
LOG.warning(msg); _append_log(job_id, msg)
time.sleep(sleep_s)
last_exc = he
except Exception as e:
sleep_s = _retry_sleep(attempt)
msg = f"{label}: error {type(e).__name__}: {e}, retry in {sleep_s:.1f}s"
LOG.warning(msg); _append_log(job_id, msg)
time.sleep(sleep_s)
last_exc = e
raise RuntimeError(f"{label}: retries exhausted: {last_exc}")
def _post_embeddings(batch: List[str], job_id: Optional[str] = None) -> Tuple[np.ndarray, int]:
last_err = None
similarity_misroute = False
for b in EMB_BACKEND_ORDER:
if b == "hf":
try:
return _call_with_retries(_hf_post_embeddings_once, batch, "HF", job_id)
except requests.HTTPError as he:
body = he.response.text if getattr(he, "response", None) is not None else ""
if "SentenceSimilarityPipeline.__call__()" in (body or ""):
similarity_misroute = True
last_err = he
_append_log(job_id, f"HF failed: {he}.")
LOG.error(f"HF failed: {he}")
elif b == "deepinfra":
try:
return _call_with_retries(_di_post_embeddings_once, batch, "DeepInfra", job_id)
except Exception as e:
last_err = e
_append_log(job_id, f"DeepInfra failed: {e}.")
LOG.error(f"DeepInfra failed: {e}")
else:
_append_log(job_id, f"Backend inconnu ignoré: {b}")
if ALLOW_DI_AUTOFALLBACK and similarity_misroute and DI_TOKEN:
LOG.warning("HF a routé sur SentenceSimilarity => auto-fallback DeepInfra (override ordre).")
_append_log(job_id, "Auto-fallback DeepInfra (HF => SentenceSimilarity).")
return _call_with_retries(_di_post_embeddings_once, batch, "DeepInfra", job_id)
raise RuntimeError(f"Tous les backends ont échoué: {last_err}")
# ---------- Chunking ----------
def _chunk_with_spans(text: str, size: int, overlap: int):
n = len(text or "")
if size <= 0:
yield (0, n, text); return
i = 0
while i < n:
j = min(n, i + size)
yield (i, j, text[i:j])
i = max(0, j - overlap)
if i >= n: break
# ---------- Background task ----------
def run_index_job(job_id: str, req: IndexRequest):
try:
_set_status(job_id, "running")
total_chunks = 0
_append_log(job_id, f"Start project={req.project_id} files={len(req.files)} | backends={EMB_BACKEND_ORDER} | store={VECTOR_STORE}")
LOG.info(f"[{job_id}] Index start project={req.project_id} files={len(req.files)}")
# Warmup -> dimension
warm = "warmup"
if req.files:
for _, _, chunk_txt in _chunk_with_spans(req.files[0].text or "", req.chunk_size, req.overlap):
if (chunk_txt or "").strip():
warm = chunk_txt; break
embs, _ = _post_embeddings([warm], job_id=job_id)
dim = embs.shape[1]
col = f"proj_{req.project_id}"
STORE.ensure_collection(col, dim)
_append_log(job_id, f"Collection ready: {col} (dim={dim})")
# loop fichiers
for fi, f in enumerate(req.files, 1):
if not (f.text or "").strip():
_append_log(job_id, f"file {fi}: vide — ignoré")
continue
batch_txts, metas = [], []
def _flush():
nonlocal batch_txts, metas, total_chunks
if not batch_txts: return
vecs, sz = _post_embeddings(batch_txts, job_id=job_id)
added = STORE.upsert(col, vecs, metas)
total_chunks += added
_append_log(job_id, f"file {fi}/{len(req.files)}: +{added} chunks (total={total_chunks})")
batch_txts, metas = [], []
for ci, (start, end, chunk_txt) in enumerate(_chunk_with_spans(f.text, req.chunk_size, req.overlap)):
if not (chunk_txt or "").strip():
continue
batch_txts.append(chunk_txt)
meta = {"path": f.path, "chunk": ci, "start": start, "end": end}
if req.store_text:
meta["text"] = chunk_txt
metas.append(meta)
if len(batch_txts) >= req.batch_size:
_flush()
_flush()
_append_log(job_id, f"Done. chunks={total_chunks}")
_set_status(job_id, "done")
LOG.info(f"[{job_id}] Index finished. chunks={total_chunks}")
except Exception as e:
LOG.exception("Index job failed")
_append_log(job_id, f"ERROR: {e}")
_set_status(job_id, "error")
# ---------- API ----------
app = FastAPI()
@app.get("/")
def root():
return {
"ok": True,
"service": "remote-indexer",
"backends": EMB_BACKEND_ORDER,
"hf_url_pipeline": HF_URL_PIPELINE if "hf" in EMB_BACKEND_ORDER else None,
"hf_url_models": HF_URL_MODELS if "hf" in EMB_BACKEND_ORDER else None,
"di_url": DI_URL if "deepinfra" in EMB_BACKEND_ORDER else None,
"di_model": DI_MODEL if "deepinfra" in EMB_BACKEND_ORDER else None,
"vector_store": VECTOR_STORE,
"vector_store_active": type(STORE).__name__,
"docs": "/health, /index, /status/{job_id}, /query, /wipe"
}
@app.get("/health")
def health():
return {"ok": True}
def _check_backend_ready():
if "hf" in EMB_BACKEND_ORDER and not HF_TOKEN:
raise HTTPException(400, "HF_API_TOKEN manquant côté serveur (backend=hf).")
if "deepinfra" in EMB_BACKEND_ORDER and not DI_TOKEN and EMB_BACKEND_ORDER == ["deepinfra"]:
raise HTTPException(400, "DEEPINFRA_API_KEY manquant côté serveur (backend=deepinfra).")
@app.post("/index")
def start_index(req: IndexRequest, background_tasks: BackgroundTasks, x_auth_token: Optional[str] = Header(default=None)):
if AUTH_TOKEN and (x_auth_token or "") != AUTH_TOKEN:
raise HTTPException(401, "Unauthorized")
_check_backend_ready()
non_empty = [f for f in req.files if (f.text or "").strip()]
if not non_empty:
raise HTTPException(422, "Aucun fichier non vide à indexer.")
req.files = non_empty
job_id = uuid.uuid4().hex[:12]
JOBS[job_id] = {"status": "queued", "logs": [], "created": time.time()}
background_tasks.add_task(run_index_job, job_id, req)
return {"job_id": job_id}
@app.get("/status/{job_id}")
def status(job_id: str, x_auth_token: Optional[str] = Header(default=None)):
if AUTH_TOKEN and (x_auth_token or "") != AUTH_TOKEN:
raise HTTPException(401, "Unauthorized")
j = JOBS.get(job_id)
if not j:
raise HTTPException(404, "job inconnu")
return {"status": j["status"], "logs": j["logs"][-800:]}
# Legacy compat
@app.get("/status")
def status_qp(job_id: str = Query(None), x_auth_token: Optional[str] = Header(default=None)):
if AUTH_TOKEN and (x_auth_token or "") != AUTH_TOKEN:
raise HTTPException(401, "Unauthorized")
if not job_id:
raise HTTPException(404, "job inconnu")
j = JOBS.get(job_id)
if not j:
raise HTTPException(404, "job inconnu")
return {"status": j["status"], "logs": j["logs"][-800:]}
class _StatusBody(BaseModel):
job_id: str
@app.post("/status")
def status_post(body: _StatusBody, x_auth_token: Optional[str] = Header(default=None)):
if AUTH_TOKEN and (x_auth_token or "") != AUTH_TOKEN:
raise HTTPException(401, "Unauthorized")
j = JOBS.get(body.job_id)
if not j:
raise HTTPException(404, "job inconnu")
return {"status": j["status"], "logs": j["logs"][-800:]}
@app.post("/query")
def query(req: QueryRequest, x_auth_token: Optional[str] = Header(default=None)):
if AUTH_TOKEN and (x_auth_token or "") != AUTH_TOKEN:
raise HTTPException(401, "Unauthorized")
_check_backend_ready()
k = int(max(1, min(50, req.top_k or 6)))
vecs, _ = _post_embeddings([req.query])
col = f"proj_{req.project_id}"
# Recherche selon le store actif
try:
hits = STORE.search(col, vecs[0], k)
except Exception as e:
raise HTTPException(400, f"Search failed: {e}")
out = []
# Qdrant renvoie des objets avec .score, .payload
for p in hits:
pl = getattr(p, "payload", None) or {}
score = float(getattr(p, "score", 0.0))
txt = pl.get("text")
if txt and len(txt) > 800:
txt = txt[:800] + "..."
out.append({
"path": pl.get("path"),
"chunk": pl.get("chunk"),
"start": pl.get("start"),
"end": pl.get("end"),
"text": txt,
"score": score,
})
return {"results": out}
@app.post("/wipe")
def wipe_collection(project_id: str, x_auth_token: Optional[str] = Header(default=None)):
if AUTH_TOKEN and (x_auth_token or "") != AUTH_TOKEN:
raise HTTPException(401, "Unauthorized")
col = f"proj_{project_id}"
try:
STORE.wipe(col); return {"ok": True}
except Exception as e:
raise HTTPException(400, f"wipe failed: {e}")
# ---------- Entrypoint ----------
if __name__ == "__main__":
import uvicorn
port = int(os.getenv("PORT", "7860"))
LOG.info(f"===== Application Startup on PORT {port} =====")
uvicorn.run(app, host="0.0.0.0", port=port)