Describe the problem

SentenceTransformerEmbeddingFunction does not expose any of the parallelism parameters that sentence-transformers natively supports. The __call__ method hard-codes only two arguments:

embeddings = self._model.encode(
    list(input),
    convert_to_numpy=True,
    normalize_embeddings=self.normalize_embeddings,
)

This means batch_size, pool, and multi-device encoding are silently unavailable. When embedding large document sets locally, this forces users to either subclass the wrapper themselves or pre-compute embeddings outside Chroma entirely — defeating the purpose of the built-in embedding function.

Describe the proposed solution

Add two optional parameters to SentenceTransformerEmbeddingFunction:

batch_size: int = 32 — passed directly to .encode(). Controls how many documents are processed per forward pass. No behaviour change for existing users.
multiprocess_devices: list[str] | None = None — when provided, starts a persistent sentence-transformers multi-process pool across the given devices (e.g. ["cuda:0", "cuda:1"] or ["cpu", "cpu", "cpu", "cpu"]). Pool is created once in __init__ and reused across every __call__ to avoid process-spawn overhead per batch.

Both parameters default to existing behaviour, so the change is fully backward compatible.

Alternatives considered

• Using FastEmbedEmbeddingFunction — FastEmbed does expose batch_size and parallel ([ENH]: FastEmbed embedding function support #1986), but it is a different library with a different model registry. It is not a drop-in replacement for users already on sentence-transformers models.
• Pre-computing embeddings outside Chroma — users can pass pre-computed embeddings directly to .add(), but this breaks the collection-level embedding function contract and requires managing embedding logic outside Chroma.

Importance

would make my life easier

Additional Information

No response