Retrieval returns 50 candidates; a cross-encoder reranker then picks the best 5. Why add this step?
Answer: B. It reads the question and each passage together, so it orders candidates more precisely than embedding similarity
Bi-encoder retrieval is fast but coarse. A cross-encoder is too slow for millions of documents but accurate on a short list, so the two are combined.