Spaces:
Sleeping
Sleeping
| """Sentence-transformer embedder for the archival tier. | |
| Lazy singleton wrapping ``SentenceTransformer``. Vectors are L2-normalized | |
| at encode time so downstream FAISS ``IndexFlatIP`` search returns cosine | |
| similarity directly (Step 16). | |
| """ | |
| from __future__ import annotations | |
| import numpy as np | |
| from sentence_transformers import SentenceTransformer | |
| from src.config import settings | |
| _MODEL: SentenceTransformer | None = None | |
| def _get_model() -> SentenceTransformer: | |
| """Load the model on first call; return the cached instance thereafter. | |
| Cache directory is set explicitly via ``cache_folder`` (not HF_HOME) so | |
| the setting scopes to just this model. On HF Spaces, the configured | |
| path lives under the persistent data dir, surviving rebuilds. | |
| """ | |
| global _MODEL | |
| if _MODEL is None: | |
| _MODEL = SentenceTransformer( | |
| settings.embeddings.model_name, | |
| cache_folder=settings.storage.embeddings_cache, | |
| ) | |
| return _MODEL | |
| def embed(text: str) -> np.ndarray: | |
| """Encode one string. Returns a normalized 1D float32 array of shape (dim,).""" | |
| model = _get_model() | |
| return model.encode(text, normalize_embeddings=True) | |
| def embed_batch(texts: list[str]) -> np.ndarray: | |
| """Encode a list of strings. Returns a normalized 2D float32 array of shape (N, dim).""" | |
| model = _get_model() | |
| return model.encode(texts, normalize_embeddings=True) | |