A second pass that reorders search results by how well they actually answer the question.
The first step of RAG retrieval casts a wide net - pulls back maybe 20-50 potentially-relevant chunks, fast. Reranking is a smarter-but-slower second pass: a specialized model reads the query AND each candidate chunk together, then scores how well that chunk actually answers the query. The top 3-5 get passed to the LLM. This two-step approach gets much better results than either step alone.
Reranking is the single fastest way to improve a mediocre RAG system. Cohere, Voyage AI, and others sell rerankers as standalone APIs. Adding a reranker usually gives a 10-30% boost in answer quality.