A team adds a reranker after vector candidate retrieval. What should the reranker primarily improve?

1. Guaranteeing that every model statement is correct
2. Ordering candidate passages by relevance to the particular question
3. Recovering documents that were never retrieved under any candidate path
4. Replacing the need for source access controls