Proposal: add rerank operation for cross-encoder / rerank-API models #253

Problem

RAG pipelines run a re-rank step after retrieval — cross-encoder or rerank-API — to score and re-order candidate documents before inference. There's no gen_ai.operation.name value for it today, so latency, cost, and per-document scores can't be attributed independently of retrieval and inference.

#217 just established gen_ai.retrieval.* as a separate namespace. rerank is the natural next operation in that direction. Distinct from gen_ai.evaluation.result (rerank produces scores, doesn't judge other models) and from embeddings (cross-encoder, no vector exposed). No existing issue mentions rerank (verified via repo search).

Proposal

Add rerank to gen_ai.operation.name + a single gen_ai.rerank.client span, with three attributes: gen_ai.rerank.candidate.count, gen_ai.rerank.top_n, gen_ai.rerank.scores (opt_in). Primary reference scenario: Cohere SDK 6.1.0 — every attribute sourced from a typed kwarg or response field. Voyage and Jina follow the same shape (Voyage uses top_k).

Open questions

1. huggingface provider enum — bundle with the rerank PR or prerequisite issue? Needed for LlamaIndex SentenceTransformerRerank.
2. Token usage shape — Cohere reports meta.billed_units.search_units (custom unit). Add gen_ai.usage.search_units opt_in, or omit usage attrs entirely until cross-provider patterns emerge?
3. gen_ai.rerank.scores cardinality — keep opt_in on the span, or downgrade to event-only?
4. Span-kind — ship client now and defer internal, or define both now?

Happy to draft the PR once SIG signals direction.