feat(litellm): align embedding instrumentation with pending spec

[codefromthecrypt]

Aligns embedding spans with semantic conventions by removing LLM-specific attributes that don't apply to embedding operations and standardizing embedding data structure:

• Removes llm.system and llm.provider from embedding spans
• Adds EMBEDDING_INVOCATION_PARAMETERS constant to semantic conventions
• Adds OPENINFERENCE_HIDE_EMBEDDINGS_VECTORS environment variable
  • Deprecates OPENINFERENCE_HIDE_EMBEDDING_VECTORS
• Adds OPENINFERENCE_HIDE_EMBEDDINGS_TEXT environment variable
  • There was no ENV before
• Updates spec documentation with rationale for attribute exclusions

Benefits:

• Clearer semantic conventions: embedding.model_name is sufficient for model identification
• Avoids ambiguity: llm.system definition unclear when applied to embeddings
• Consistent with established patterns: embedding spans use embedding.* prefix exclusively

Breaking changes:

openinference-instrumentation-openai:

• llm.system attribute removed from embedding spans
• llm.provider attribute removed from embedding spans

openinference-instrumentation-litellm:

• Attribute names change from embedding.text → embedding.embeddings.{i}.embedding.text
• Attribute names change from embedding.vector → embedding.embeddings.{i}.embedding.vector
• embedding.vector now captures all embeddings (not just first) as tuples (not JSON strings)

openinference-instrumentation (TraceConfig):

• hide_embedding_vectors behavior changed: now returns "__REDACTED__" instead of removing attribute (None)
  • This aligns with documented spec behavior for redacted content

Deprecations:

openinference-instrumentation (TraceConfig):

• OPENINFERENCE_HIDE_EMBEDDING_VECTORS deprecated → use OPENINFERENCE_HIDE_EMBEDDINGS_VECTORS
• hide_embedding_vectors deprecated → use hide_embeddings_vectors
• Both old and new continue to work via OR logic for backwards compatibility

---

Note

Standardizes embedding spans across SDKs: use CreateEmbeddings, embed invocation params and per-item text/vector, drop LLM system/provider for embeddings, and add new embedding redaction config; updates specs and tests accordingly.

• Embedding Semantics & Specs:
  • Add SpanAttributes.EMBEDDING_INVOCATION_PARAMETERS; update docs to use span name CreateEmbeddings, record per-item embedding.embeddings.{i}.embedding.text/vector, and clarify no llm.system/provider for embeddings.
  • Revise spec for text vs token inputs and vector handling; update configuration docs with new redaction env vars and deprecations.
• Config (TraceConfig):
  • Add OPENINFERENCE_HIDE_EMBEDDINGS_VECTORS (deprecates OPENINFERENCE_HIDE_EMBEDDING_VECTORS) and OPENINFERENCE_HIDE_EMBEDDINGS_TEXT; masking now uses "__REDACTED__".
• OpenAI Instrumentation:
  • Embedding spans: name CreateEmbeddings, emit embedding.invocation_parameters, exclude llm.system/provider.
• LiteLLM Instrumentation:
  • Embeddings: name CreateEmbeddings; capture invocation params; record input texts and all vectors under embedding.embeddings.{i}.*; set embedding.model_name from response when present.
• Tests & Cassettes:
  • Add embedding tests (single/batch, model name override), update assertions to new keys and span names; adjust token count/instrumentation fixtures.

^{Written by Cursor Bugbot for commit b9dc097. This will update automatically on new commits. Configure here.}

[codefromthecrypt]

so this is the same as #2210, and a FAQ might be why return an indexed list? The reason is that if you have a mixed batch request, you'll get multiple embedding vectors back, index correlated with the input text. one example recorded from the openai instrumentation is this one https://github.com/envoyproxy/ai-gateway/blob/main/tests/internal/testopeninference/spans/embeddings-mixed-batch.json

[codefromthecrypt]

@axiomofjoy I updated this PR including the specs along with your rationale about why embeddings spans should have no system/provider attributes. If I got anything wrong, lemme know and I'll bump it!

[codefromthecrypt]

#2295 for the unrelated inspector test failures

[axiomofjoy]

Was this pattern causing an issue?

[axiomofjoy]

Thanks for taking a stab at this. I'm not sure we need to proscribe using llm.system and llm.provider for embedding spans. Only that we've never used these attributes to describe a client interface like litellm before. In the case of the OpenAI SDK, I think it could make sense to set them. It might make less sense to set them if we don't know the specific API being hit, as is the case for LiteLLM.

Let me chat with @RogerHYang to see if he has the same understanding.