A second pass that reorders search results before they reach the model. Fast similarity search is good at getting roughly the right twenty documents and poor at deciding which three are best. A reranker examines each candidate against the actual question more carefully, and it typically improves answer quality more than any other single change to a retrieval system.
Technical view
Cross-encoders score query and document jointly rather than comparing independently computed embeddings, which is far more accurate and far too slow to run across a whole corpus, hence the retrieve-then-rerank pattern. Typical configurations retrieve fifty to a hundred candidates and rerank down to three to five. Hybrid retrieval combining sparse keyword matching with dense vectors, fused by reciprocal rank fusion, is a common companion, since keyword search still wins on names, codes and rare terms.
Related
Part of AI Terminology Index, section: Retrieval and Knowledge.