Fix(embeddings): Wrap each separate input in a Content+Part to fix batching

[yorickvP]

gemini-embedding-2-preview was interpreting an array as a single multi-part embedding request, causing only one embedding to be returned.

This commit wraps each separate input in a Content+Part object to fix that issue.

• Closes Bug when using gemini-embedding-2-preview #4872

Pre-Review Checklist

• Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
• No breaking changes in accordance with the version policy.
• Linting and type checking pass per make format and make typecheck.
• PR title is fit for the release changelog.

Pre-Merge Checklist

• New tests for any fix or new behavior, maintaining 100% coverage.
• Updated documentation for new features and behaviors, including docstrings for API docs.

[devin-ai-integration]

Devin Review found 1 potential issue.

[devin-ai-integration]

🚩 Content role defaults to None instead of 'user'

The old code passed raw strings (list[str]) which the Google SDK internally converted to Content objects with role='user'. The new code at line 166 creates Content(parts=[Part(text=text)]) without specifying role, which defaults to role=None. I verified this by inspecting the SDK: Content(parts=[Part(text='hello')]) yields role=None.

The existing VCR cassette (tests/cassettes/test_embeddings/TestGoogle.test_query.yaml:22) shows role: user in the recorded request body. However, VCR matching is configured in tests/conftest.py to only match on method and path (not body), so tests still pass.

For the embedding API specifically, the role field is semantically irrelevant — the API extracts text from parts regardless of role. This is not a bug, but if strict request parity with the old behavior is desired, role='user' could be added explicitly. The cassettes should ideally be re-recorded to reflect the actual new request format.

---

Was this helpful? React with 👍 or 👎 to provide feedback.