[NishchayMahor]

What

The async embedding helpers validate input_type but then drop it. In voyageai/embeddings_utils.py:

• Sync _get_embeddings forwards it: Embedding.create(input=..., model=..., input_type=input_type, **kwargs).
• Async _aget_embeddings validates it (_check_input_type(input_type)) but calls Embedding.acreate(input=..., model=..., **kwargs) — input_type is a named parameter, not part of **kwargs, so it never reaches the API.

Both public async functions (aget_embedding, aget_embeddings) route through _aget_embeddings, so:

await aget_embedding(text, model="voyage-3", input_type="query")

computes the embedding with no input-type prefix, which degrades retrieval quality — a silent correctness bug and a sync/async divergence.

Fix

Forward input_type in the async call, mirroring the sync path:

await voyageai.Embedding.acreate(
    input=list_of_text, model=model, input_type=input_type, **kwargs
)

Testing

Added test_aget_embeddings_forwards_input_type to TestEmbeddingsUtils (mirrors the existing test_aget_embeddings_internal mocked pattern) asserting acreate is called with input_type="query". Fully offline (mocks acreate, no API key). It fails on main and passes with the fix.

[fzowl]

Verified locally with a live API key: full suite 370 passed / 21 skipped (local-model tests skip without local extras); the only failure was a transient APIConnectionError on an async multimodal case that moved between parametrizations across runs — environmental network flakiness, unrelated to this change. All embedding tests (the code path touched here) pass.

The fix is correct: async _aget_embeddings now forwards input_type to acreate, mirroring the sync path; since input_type is a named parameter it was previously silently dropped. No double-pass risk. The added offline test asserts the forwarded kwarg.

Note: the fork PR's Integration/Tests CI checks are red only because fork PRs don't receive the VOYAGE_API_KEY secret; with a key they pass (verified above).

[fzowl]

@NishchayMahor Thank you very much for your contribution, great catch! I'm releasing a new version with this fix very soon!

[NishchayMahor]

Thanks for the merge! Small fix, but the quiet kind that bites people at exactly the wrong time, so glad it's in. Big fan of the embedding models, retrieval quality is something we lean on a lot in the observability work I do. Happy to help out here again anytime.