fix(gemini): filter params from embedding requests

[Chesars]

Relevant issues

Fixes #24293

Pre-Submission checklist

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Type

🐛 Bug Fix

Changes

The Gemini batch embedding transformation spreads all optional_params into the request body via **gemini_params. Params like max_tokens (injected by add_provider_specific_params_to_optional_params for non-OpenAI providers) reach the Gemini API and cause a 400 BadRequestError.

drop_params: true doesn't prevent this because the param is re-injected after the drop_params check runs.

Extract _filter_embed_params() that maps dimensions/task_type and keeps only the fields Gemini embeddings accept (outputDimensionality, taskType, title). Applied to both transform_openai_input_gemini_content and transform_openai_input_gemini_embed_content.

Tests added

• test_filter_embed_params_drops_unsupported — verifies max_tokens/temperature are filtered
• test_filter_embed_params_keeps_supported — verifies dimensions/task_type/title pass through
• test_batch_embed_content_drops_max_tokens — integration test for batchEmbedContents path
• test_embed_content_drops_max_tokens — integration test for embedContent path

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 22, 2026 8:43pm

[greptile-apps]

Greptile Summary

This PR fixes a 400 BadRequestError on Gemini embedding requests by introducing a _filter_embed_params() helper that whitelists only the three fields Gemini embeddings accept (outputDimensionality, taskType, title) and applies it to both the batchEmbedContents and embedContent transformation paths.

• Root cause addressed: add_provider_specific_params_to_optional_params injects max_tokens (and potentially other non-embedding params) into optional_params after the drop_params check runs; previously, these were blindly spread into every Gemini embedding request body via **gemini_params.
• Fix: _filter_embed_params() centralises param mapping (dimensions → outputDimensionality, task_type → taskType) and strict whitelist filtering in one place, replacing duplicated inline logic in both transformation functions.
• Tests: Four new pure unit tests are added — two for the helper function directly and two integration tests through the transformation functions — none of which make real network calls, satisfying the mock-only test policy for this directory.
• Type safety: The whitelist ({"outputDimensionality", "taskType", "title"}) correctly matches the optional fields declared in the VertexAITextEmbeddingsRequestBody TypedDict that EmbedContentRequest inherits from.

Confidence Score: 5/5

• This PR is safe to merge — it fixes a targeted 400 error with no backwards-incompatible surface changes.
• The change is minimal and surgical: it extracts existing inline mapping logic into a shared helper and adds a whitelist filter. The whitelist exactly matches the optional fields of the EmbedContentRequest / VertexAITextEmbeddingsRequestBody TypedDict, so no valid params are dropped. All four new tests are pure unit/mock tests, coverage for both endpoints is present, and no existing tests are weakened.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/vertex_ai/gemini_embeddings/batch_embed_content_transformation.py	Extracts a _filter_embed_params() helper that maps OpenAI field names (dimensions → outputDimensionality, task_type → taskType) and whitelists only the three fields Gemini embeddings actually accept, preventing unsupported params like max_tokens from being spread into the request body.
tests/litellm/llms/vertex_ai/test_gemini_batch_embeddings.py	Adds four focused unit tests for the new _filter_embed_params helper and integration tests for both embedding endpoints; all tests are pure mock/unit tests with no real network calls.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["litellm.embedding() called\nwith optional_params\ne.g. {max_tokens: 256, dimensions: 768}"] --> B["_filter_embed_params(optional_params)"]
    B --> C{"Map OpenAI keys"}
    C -->|"dimensions → outputDimensionality"| D["Rename key"]
    C -->|"task_type → taskType"| E["Rename key"]
    C -->|other keys unchanged| F["Keep as-is"]
    D & E & F --> G["Whitelist filter\n_SUPPORTED_EMBED_PARAMS\n= {outputDimensionality, taskType, title}"]
    G -->|"key in whitelist"| H["Include in gemini_params"]
    G -->|"key NOT in whitelist\ne.g. max_tokens, temperature"| I["Drop param ❌"]
    H --> J{"Which endpoint?"}
    J -->|"Text input"| K["transform_openai_input_gemini_content\n→ batchEmbedContents"]
    J -->|"Multimodal input"| L["transform_openai_input_gemini_embed_content\n→ embedContent"]
    K --> M["EmbedContentRequest(**gemini_params)\nNo 400 BadRequestError"]
    L --> N["request_body dict(**gemini_params)\nNo 400 BadRequestError"]

Loading

_{Reviews (1): Last reviewed commit: "fix(gemini): filter unsupported params f..." | Re-trigger Greptile}

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing Chesars:fix/gemini-embedding-drop-unsupported-params (db0d85e) with main (c89496f)}