A second-pass model that reorders retrieved chunks by true relevance to the query.
1 min readupdated 2026-06-22
/ quick answer
In RAG, vector search returns a candidate set that's fast but noisy. A reranker (usually a cross-encoder like Cohere Rerank or bge-reranker) scores each candidate against the query and reorders them, dramatically improving answer quality. A second-pass model that reorders retrieved chunks by true relevance to the query.
A second-pass model that reorders retrieved chunks by true relevance to the query. In RAG, vector search returns a candidate set that's fast but noisy. A reranker (usually a cross-encoder like Cohere Rerank or bge-reranker) scores each candidate against the query and reorders them, dramatically improving answer quality. In practice: Retrieve top 50 chunks by vector similarity → rerank down to top 5 → send only those to the LLM. Latency +200ms, accuracy often +20–40%. This dictionary node is part of the Onexial knowledge graph and links to related concepts, workflows and tools below.
Definition
In RAG, vector search returns a candidate set that's fast but noisy. A reranker (usually a cross-encoder like Cohere Rerank or bge-reranker) scores each candidate against the query and reorders them, dramatically improving answer quality.
Example
Retrieve top 50 chunks by vector similarity → rerank down to top 5 → send only those to the LLM. Latency +200ms, accuracy often +20–40%.
In RAG, vector search returns a candidate set that's fast but noisy. A reranker (usually a cross-encoder like Cohere Rerank or bge-reranker) scores each candidate against the query and reorders them, dramatically improving answer quality.
What is an example of Reranking?
Retrieve top 50 chunks by vector similarity → rerank down to top 5 → send only those to the LLM. Latency +200ms, accuracy often +20–40%.
Why does Reranking matter for AI and automation?
A second-pass model that reorders retrieved chunks by true relevance to the query. It connects to the workflows, prompts and tool stacks linked on this page, so you can move from definition to execution without leaving Onexial.