Principle:Cohere ai Cohere python Semantic Reranking
| Field | Value |
|---|---|
| Type | Principle |
| Source | Cohere Python SDK |
| Domain | NLP Information Retrieval Reranking |
| Last Updated | 2026-02-15 |
| Implemented By | Implementation:Cohere_ai_Cohere_python_V2Client_Rerank |
Overview
A technique for re-ordering candidate documents by semantic relevance to a query using cross-encoder models.
Description
Semantic Reranking uses a cross-encoder model to score the relevance of each candidate document against a query. Unlike embedding-based retrieval (which uses separate encoders for query and document), reranking jointly attends to both query and document text, producing more accurate relevance scores at the cost of higher computation. This is typically used as a second stage in a retrieval pipeline: initial retrieval (fast, approximate) produces candidates, then reranking (slower, precise) re-orders them.
Usage
Use as the second stage in a two-stage retrieval pipeline. After initial vector similarity search returns top-K candidates, pass the query and candidate texts to the rerank API to get precision-ordered results.
Theoretical Basis
Cross-encoder reranking processes query-document pairs jointly through a transformer, computing attention between all tokens. This is more accurate than bi-encoder similarity but O(N) per document rather than O(1) with pre-computed embeddings. The relevance_score is a normalized probability (0-1) from the model's relevance head.