What is reranking in search and RAG?

Definition: Reranking is a second retrieval step that takes the top results from a fast search and reorders them with a more accurate model, so the most relevant passages reach the user or the language model first.

How reranking works

A first-stage retriever, such as vector search, keyword search or a hybrid of both, quickly returns a few dozen candidates. A reranker, often a cross-encoder model that reads the query and each passage together, scores every candidate for relevance. The system keeps the top few for the final answer.

Why it matters

First-stage search is tuned for speed and recall, so relevant passages can land lower in the list. In RAG, putting the right passages first means the model sees better context and the prompt stays shorter. Rerankers add latency, so teams limit how many candidates they score.

How teams use it

Teams add a hosted or open-weight reranker on top of semantic search, then measure the gain with labeled test queries before and after the change, tracking metrics such as recall and mean reciprocal rank.

How Ryz helps

Ryz RAG engineers tune retrieval and reranking in your cloud, working on your team and reporting to your leads.

Related

• RAG development services
• Hire RAG engineers
• Semantic search
• Vector database