"""Local BGE embeddings over authorized chunks; explicit failure without a model.""" import os import threading from pathlib import Path from APIGateway.app.errors import GatewayError _lock = threading.RLock() _model = None _vectors = {} MODEL_ID = 'BAAI/bge-small-zh-v1.5' def model_path(): configured = os.getenv('WORKOS_EMBEDDING_MODEL_PATH') if configured: return configured homes = [Path(os.getenv('HF_HOME', str(Path.home()/'.cache/huggingface'))), Path(r'E:\codex\models\huggingface'), Path(r'D:\codex\models\huggingface')] for home in homes: for folder in sorted((home/'hub/models--BAAI--bge-small-zh-v1.5/snapshots').glob('*')): if (folder/'model.safetensors').is_file():return str(folder) return MODEL_ID def rank_semantic(query, rows, top_k): if not rows:return [] global _model with _lock: try: import numpy as np if _model is None: from sentence_transformers import SentenceTransformer _model = SentenceTransformer(model_path(), device='cpu', local_files_only=True) missing=[(c,d) for c,d in rows if c.id not in _vectors] if missing: vectors=_model.encode([c.text for c,d in missing], batch_size=32, normalize_embeddings=True, show_progress_bar=False) _vectors.update({c.id:v for (c,d),v in zip(missing,vectors)}) if len(_vectors)>10000: keep={c.id for c,d in rows};_vectors.update({}) for key in list(_vectors): if key not in keep:del _vectors[key] vector=_model.encode(['为这个句子生成表示以用于检索相关文章:'+query], normalize_embeddings=True, show_progress_bar=False)[0] scores=np.asarray([_vectors[c.id] for c,d in rows]) @ vector except (ImportError, OSError, RuntimeError) as exc: raise GatewayError(503, 'EMBEDDING_UNAVAILABLE', '本地 BGE 向量模型未准备好;请安装向量依赖并配置 WORKOS_EMBEDDING_MODEL_PATH。') from exc scored=[(float(score),c,d) for score,(c,d) in zip(scores,rows)] return sorted(scored,key=lambda x:(-x[0],x[1].id))[:top_k]