Reranking

First-stage search, whether keyword or embedding based, is built for speed across a large collection, which means its ordering is approximate. Reranking takes the top results, often a few dozen, and scores each one against the query more carefully.

Rerankers are usually cross-encoder models that read the query and a candidate passage together, which captures relevance better than comparing two separately computed embeddings. A language model can also act as a reranker for small result sets.

The payoff is that you can pass fewer, better passages to the model, which improves answers and saves context. The cost is extra latency and another component to run, so measure whether it improves your retrieval evals before adding it.