What is Reranking? - Definition & Meaning

Definition

Reranking is the reordering of retrieved documents with a more accurate model (often a cross-encoder) to place the most relevant results at the top. It significantly improves RAG quality compared to vector search alone.

Technical explanation

First, many candidates (e.g., 100) are fetched via fast vector or keyword search. Then a cross-encoder scores query-document pairs and reorders by relevance. Cross-encoders are more accurate than bi-encoders (embedding similarity) but slower. Cohere Rerank, Jina Reranker, and open source (ms-marco, BGE-reranker) are common. Trade-off: more candidates = better recall but higher latency.

How AVARC Solutions applies this

AVARC Solutions adds reranking to retrieval pipelines where quality is critical. We use Cohere Rerank or open source cross-encoders. For high-traffic systems we limit reranking to the top 20–30 candidates.

Practical examples

• A RAG fetching 50 chunks via vector search, then selecting the top 5 with Cohere Rerank for the LLM.
• An enterprise search combining hybrid retrieval with a reranker for the most accurate results.
• A support chatbot where reranking ensures the right FAQ chunks appear at the top.

Related terms