Reranking reorders retrieved candidates using a stronger relevance model.

Retrieval often gets a broad candidate set quickly, while reranking improves precision before context is passed to the generator.