fix(openai): filter None values from embedding optional_params (encoding_format=null)

[IvoryKir]

Summary

Filter out None and empty string values from optional_params in the generic OpenAI embedding handler before constructing the request, preventing encoding_format: null from being sent to custom endpoints.

Problem

When using litellm.embedding() with the openai provider and a custom api_base pointing to a local embedding server (llama.cpp, vLLM, TEI, etc.), the request includes encoding_format: null in the JSON body. Strict JSON parsers reject this:

[json.exception.type_error.302] type must be string, but is null

This was already fixed for the openai_like handler (L108-110 in litellm/llms/openai_like/embedding/handler.py) after the vLLM embeddings incident, but the same filter was not applied to the generic openai handler in litellm/llms/openai/openai.py.

Users hitting this issue use the openai provider with api_base to point to local servers — a common pattern for llama.cpp, TEI, LocalAI, etc.

Fix

Apply the exact same filtering pattern from openai_like handler to the generic openai handler:

# Before (L1327):
data = {"model": model, "input": input, **optional_params}

# After:
filtered_optional_params = {
    k: v for k, v in optional_params.items() if v not in (None, "")
}
data = {"model": model, "input": input, **filtered_optional_params}

Impact

• No behavioral change for standard OpenAI API calls (they don't send None values)
• Fixes all custom api_base endpoints that reject null JSON values
• Consistent with the openai_like handler's existing approach

Refs

• vLLM incident report: https://docs.litellm.ai/blog/vllm-embeddings-incident
• Related issue: [Bug]: Embedding busted on 1.80.11 #19174
• openai_like fix (reference): litellm/llms/openai_like/embedding/handler.py L108-110

[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 7:55am

[CLAassistant]

All committers have signed the CLA.

[codspeed-hq]

Merging this PR will not alter performance

✅ 16 untouched benchmarks

---

_{Comparing IvoryKir:fix/openai-embedding-encoding-format-null (a2386dd) with main (c89496f)}

[greptile-apps]

Greptile Summary

This PR applies the same None/empty-string filtering already present in the openai_like embedding handler to the generic openai embedding handler, preventing encoding_format: null (and any other None-valued optional params) from being serialised into the JSON body sent to strict OpenAI-compatible endpoints such as llama.cpp, vLLM, and TEI.

Changes:

• litellm/llms/openai/openai.py — one-line dict comprehension filter added before the data dict is constructed; the filtered dict is forwarded correctly to both the sync (make_sync_openai_embedding_request) and async (aembedding) paths.
• New test file with 6 mock-based cases covering None, "", "float", "base64", mixed params, and empty optional_params. No real network calls are made, satisfying CI requirements.

Minor observation: The filter (v not in (None, "")) silently drops any empty-string optional param, not only encoding_format. This is intentional and exactly mirrors the openai_like handler's behaviour, but it is worth noting that a caller explicitly passing e.g. user="" would have that param silently omitted. This is unlikely to be a practical concern given OpenAI's API contracts.

Confidence Score: 4/5

• Safe to merge — the change is a minimal, targeted fix that is consistent with the existing openai_like handler pattern and has comprehensive mock-based test coverage.
• The fix is a single dict comprehension that mirrors an identical, battle-tested pattern already in the codebase. It correctly covers both the sync and async code paths (async aembedding receives the already-filtered data dict). Six mock-based tests cover all meaningful cases. No real network calls in the test suite. Confidence is 4 rather than 5 because the filter is applied to all optional params (not just encoding_format), which is intentional but slightly broader than the stated scope.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/openai/openai.py	Adds a filter to remove None and "" values from optional_params before building the embedding request data dict, mirroring the pattern already in the openai_like handler. The filtered data dict is correctly passed to both the sync and async paths.
tests/test_litellm/llms/openai/embedding/test_openai_embedding_encoding_format.py	New mock-only test file covering 6 cases (None, empty string, float, base64, other params, empty dict). Uses patch.object on make_sync_openai_embedding_request and _get_openai_client, and also patches convert_to_model_response_object to prevent false failures on unrelated validation. No real network calls are made.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["OpenAIChatCompletion.embedding()"] --> B["Build filtered_optional_params\n{k: v for k, v in optional_params.items()\n if v not in (None, '')}"]
    B --> C["data = {model, input, **filtered_optional_params}"]
    C --> D{aembedding?}
    D -- "True (async)" --> E["aembedding(data=data, ...)"]
    D -- "False (sync)" --> F["_get_openai_client()"]
    F --> G["make_sync_openai_embedding_request(data=data, ...)"]
    G --> H["convert_to_model_response_object()"]
    E --> H
    H --> I["Return EmbeddingResponse"]

    style B fill:#c8f7c5,stroke:#27ae60
    style C fill:#c8f7c5,stroke:#27ae60

Loading

_{Last reviewed commit: "fix(tests): mock con..."}

[IvoryKir]

The lint CI failure is unrelated to this PR — Black flags litellm/proxy/management_helpers/audit_logs.py which is not touched by these changes. This is a pre-existing formatting issue in the main branch.