Reranking uses Cross-Encoder models to perform fine-grained scoring and reordering of candidate documents against the query, improving the quality of context sent to the LLM.