fix(embedding): add embed_query/embed_documents to EmbeddinggemmaONNX for ChromaDB 1.5.x

[ghost]

Bug

ChromaDB 1.5.x calls embedding_function.embed_query(input=...) via keyword argument during collection.query(). EmbeddinggemmaONNX lacked both embed_query and embed_documents methods, causing:

TypeError: embed_query() got an unexpected keyword argument input

whenever semantic search was triggered.

Fix

This PR adds the two methods required by the ChromaDB EF protocol, using input (the ChromaDB kwarg name, noqa A002) so that palace search works correctly with the embeddinggemma model.

Also downloads the companion _data ONNX file alongside the main model to prevent runtime InferenceSession failures.

Related

Fixes silent search failures when embedding_model is set to "embeddinggemma".

[gemini-code-assist]

Code Review

This pull request updates mempalace/embedding.py to download the ONNX model data file during lazy loading and adds embed_query and embed_documents methods to support the ChromaDB embedding function protocol. A critical issue was identified in embed_query: if it is called with a single string (as is common in integrations like LangChain), it will iterate over individual characters instead of treating the string as a single input. A suggestion was provided to handle both string and list inputs correctly.

[gemini-code-assist]

If embed_query is called with a single string (which is the standard signature for embed_query in LangChain and other common integrations), input will be a str instead of a list[str].

Since EmbeddinggemmaONNX.__call__ expects an iterable of strings, passing a single string will cause Python to iterate over its individual characters.

Suggested change

	def embed_query(self, input: list[str]) -> list[list[float]]: # noqa: A002 — ChromaDB EF protocol
	"""Embed query documents (ChromaDB EF protocol)."""
	return self(input)
	def embed_query(self, input: str | list[str]) -> list[float] | list[list[float]]: # noqa: A002 — ChromaDB EF protocol
	"""Embed query documents (ChromaDB EF protocol)."""
	if isinstance(input, str):
	return self([input])[0]
	return self(input)

[igorls]

Maintainer push: added one commit (48f4eb6) updating test_lazy_load_runs_once to expect 3 hf_hub_download calls instead of 2. Your fix correctly adds the model.onnx_data external-weights download (the model can't load weights without it), so a single warm-up now legitimately issues 3 downloads. The lazy-load-once invariant still holds — InferenceSession and Tokenizer.from_file are each still built exactly once. Full test_embeddinggemma.py is green locally (10 passed). The embed_query/embed_documents EF-protocol additions look correct for ChromaDB 1.5.x. Thanks!

[mvalentsev]

#1667 independently adds the same model_quantized.onnx_data sibling fetch (sidecar only). Flagging the overlap for dedup. This PR also bumps test_lazy_load_runs_once for the extra hf_hub_download and adds the embed_query/embed_documents methods, which #1667 doesn't.