D
Doc and Tell
Glossary/ai/ml
ai/ml

Reranking

A second-stage retrieval step that re-scores initially retrieved documents using a more powerful model to improve the relevance of the final results.

Initial retrieval (vector search + BM25) is optimized for speed and recall — it retrieves a broad set of potentially relevant passages. Reranking applies a cross-encoder model that jointly processes the query and each candidate passage to produce a relevance score that captures more nuanced semantic relationships than the bi-encoder approach used for initial retrieval. The top-k passages after reranking are then passed to the LLM for answer generation.

Reranking improves answer quality significantly for complex, multi-part questions where the initial retrieval may surface passages that are topically related but not specifically responsive to the query. Cross-encoder models like Cohere Rerank, BGE Reranker, and ColBERT achieve substantially higher precision than bi-encoder retrieval alone, at the cost of higher latency and API fees. For professional document intelligence applications in legal and compliance contexts — where missing a critical passage can have real consequences — the quality improvement typically justifies the additional cost.

Analyze Documents Related to Reranking

Upload any document and get AI-powered analysis with verifiable citations.

Start Free