reranker

rag-reranker

Two-stage cross-encoder reranker + MMR diversity for AgentContextOS.

Step 2.7 deliverable. Anchored in ADR-0006.

What ships

Component	Where	Notes
RerankPipeline	rag_reranker.pipeline	Orchestrator: fast_rerank → optional early-exit → hydrate → precise_rerank → optional MMR. Emits the canonical rerank OTel span.
mmr_rerank	rag_reranker.mmr	Pure Carbonell–Goldstein MMR primitive.
SentenceTransformersCrossEncoder	rag_reranker.backends.sentence_transformers	Local cross-encoder; default factory bge_reranker_v2_m3(), fast factory minilm_l6().
CohereReranker	rag_reranker.backends.cohere	Cohere Rerank v3 API; precise stage only.
JinaReranker	rag_reranker.backends.jina	Jina Reranker hosted API; precise stage only.

Install

# Core (works against the noop reranker out of the box)
uv add rag-reranker

# Local cross-encoder (BAAI/bge-reranker-v2-m3 or MiniLM)
uv add 'rag-reranker[sentence-transformers]'

# Hosted APIs
uv add 'rag-reranker[cohere]'
uv add 'rag-reranker[jina]'

See also

• docs/reference/reranker.md — public API + extension points
• docs/architecture/reranking.md — two-stage rationale + MMR algorithm + composition with HybridRetriever
• docs/adr/ADR-0006-two-stage-rerank.md — the design decision