LiteLLM: `patch_litellm()` does not patch `embedding` or `moderation`; `aembedding` missing entirely

[braintrust-bot]

Gap

patch_litellm() — the function used by auto_instrument() — only replaces 4 of the 6 functions that LiteLLMWrapper instruments. Embedding and moderation calls made via litellm.embedding() or litellm.moderation() are silently untraced when using auto-instrumentation.

Additionally, LiteLLMWrapper itself is missing aembedding() (async embeddings), creating an asymmetry with the rest of the wrapper which provides both sync and async variants for completion and responses.

patch_litellm() (line 630–667) patches:

Function	Patched?
litellm.completion	Yes
litellm.acompletion	Yes
litellm.responses	Yes
litellm.aresponses	Yes
litellm.embedding	No
litellm.aembedding	No (also missing from LiteLLMWrapper)
litellm.moderation	No
litellm.amoderation	No (also missing from LiteLLMWrapper)

LiteLLMWrapper (line 484–518) exposes:

• completion() / acompletion() — sync + async
• responses() / aresponses() — sync + async
• embedding() — sync only, no async
• moderation() — sync only, no async

What is missing

1. patch_litellm() should also patch litellm.embedding and litellm.moderation so auto-instrument users get tracing on these calls.
2. LiteLLMWrapper should add aembedding() and amoderation() for async parity, and patch_litellm() should patch those as well.

Embeddings are a core AI API surface used in RAG pipelines and semantic search. The EmbeddingWrapper class (line 423) already exists and works correctly — it's just not wired into patch_litellm().

Braintrust docs status

not_found — no LiteLLM-specific integration page was found at braintrust.dev/docs/integrations/ai-providers/litellm (404).

Upstream sources

• LiteLLM embedding docs: https://docs.litellm.ai/docs/embedding/supported_embedding
• LiteLLM aembedding() function: https://docs.litellm.ai/docs/embedding/async_embedding
• LiteLLM moderation: https://docs.litellm.ai/docs/moderation

Local files inspected

• py/src/braintrust/wrappers/litellm.py:
  • LiteLLMWrapper.__init__ (line 484–494) — creates _embedding_wrapper and _moderation_wrapper but no async variants
  • LiteLLMWrapper class (line 496–518) — exposes embedding() and moderation() sync only
  • patch_litellm() (line 630–667) — only patches completion, acompletion, responses, aresponses
  • EmbeddingWrapper class (line 423) — fully implemented with metrics extraction
  • TODO at line 446: "Add a flag to control whether to log the full embedding vector"
• py/src/braintrust/wrappers/test_litellm.py — has embedding tests but only via wrap_litellm(), not via patch_litellm()
• Grep for aembedding in litellm.py: zero results