A second pass that re-scores retrieved candidates by reading each one against the query, fixing the ordering that fast vector search got approximately right.
Vector search compares a query embedding to document embeddings computed independently, which is fast and imprecise. A reranker reads the query and the passage together and scores the pair directly, which is slow and much more accurate — so it runs over the top fifty candidates rather than the whole corpus.
It is usually the cheapest large improvement available to a retrieval pipeline. If the right passage is somewhere in your top fifty but never in the top five, reranking is the fix, not a better embedding model.