Reranker

A reranker is a model that takes the top-K results from a retrieval step and reorders them by relevance to the query — improving RAG quality by fixing the ordering errors of pure embedding search.

Embedding search finds candidates; rerankers pick winners. A typical stack: retrieve top-50 by embeddings (fast, broad), rerank to top-5 with a cross-encoder (slow, accurate), pass top-5 to the LLM. Adding a reranker is the highest-impact RAG quality lever in 2026.

Example

A RAG system retrieves 50 candidates with cosine similarity, reranks them with Cohere Rerank or BGE-reranker, and ends up with 5 chunks the LLM can confidently cite.

How OpenLabor uses it

OpenLabor's RAG pipeline uses a reranker by default on knowledge-heavy missions.

Are rerankers expensive?

Cheap relative to LLM calls. Reranking 50 chunks costs cents; the quality lift is large.

Related: rag, embedding, vector-store, hybrid-search.

AI Labor Glossary