rag/embedder.py

import os

from langchain_huggingface import HuggingFaceEmbeddings

_embedder = None


def get_embedder() -> HuggingFaceEmbeddings:
    """
    Local HuggingFace embedding model — BAAI/bge-small-en-v1.5
    Downloaded once to ~/.cache/huggingface on first run.
    normalize_embeddings=True required for correct cosine similarity with BGE.
    """
    global _embedder
    if _embedder is None:
        _embedder = HuggingFaceEmbeddings(
            model_name=os.getenv("HF_EMBEDDING_MODEL", "BAAI/bge-small-en-v1.5"),
            model_kwargs={"device": "cpu"},
            encode_kwargs={"normalize_embeddings": True},
        )
    return _embedder