Text Reranking
Score and reorder candidate texts by relevance to a query
Reranking models are cross-encoders that jointly score a query and a candidate text, producing a relevance score. Use as a second-stage step after fast embedding-based retrieval to significantly improve ranking quality.
Typical pipeline:
- Retrieve top-100 candidates using a fast embedding model
- Rerank with a cross-encoder to get the best top-10
Available Models
- BGE Reranker V2 Large – Open-weight multilingual cross-encoder, top performance on MS MARCO and BEIR
- Cohere Rerank v3 – Commercial-grade reranker for enterprise search and retrieval