A reranker model receives a query and candidate text, then estimates which candidates are most relevant. Unlike a first-stage retriever optimized to search a large index quickly, a reranker can spend more computation comparing each candidate with the query. It is often used to improve results in search and retrieval-augmented generation.
Developers often compare a reranker's top results against a labeled set of queries, checking whether a known good document lands in the first three. Hosted reranking services and open models both exist, so you can start with a hosted option and move later if cost or privacy requires it. Check the language and domain the model was trained on, since a reranker built for English web text may do worse on legal or code content. First-stage candidates usually come from embeddings in a vector database.
Choose a candidate pool that balances recall, latency and cost. Evaluate ranking quality with known relevant results, and remember that a reranker cannot add missing candidates to the list it receives. Related: Search Reranking, Hybrid Search.