Reranking is the polish step in a RAG pipeline: after retrieval gathers candidate passages, a reranker scores each one against the actual question and reorders them so the best answer is first.
Why retrieval order isn't good enough
Embedding similarity is computed once, in the abstract — it doesn't look at the specific question and passage together. A cross-encoder reranker does: it reads the query and each candidate as a pair and produces a sharp relevance score, catching subtle mismatches that vector distance misses.
The payoff: pass fewer, better chunks
With a good reranker, the top 3 passages are usually enough — so you send less context, spend fewer tokens, and get answers with less noise. Reranking is often the highest-leverage upgrade to an existing RAG system.
Where it fits
Reranking sits right after hybrid search and right before generation. It's also what makes RAG evaluation move — reorder quality is measurable. Kit for AI reranks in-house on every retrieval, no extra service to run. Start free to try it.