The idea, in plain words
Score the question and passage together
The first retriever quickly selects possible passages. A cross-encoder then reads the question and each candidate together to score their relevance more closely.
This lesson uses recorded real model outputs. Increasing the candidate pool gives the reranker more passages to inspect, but requires more query–passage scoring work.
1
Take only the candidate pool produced by the first retrieval stage.
2
Pair the original question with each candidate passage.
3
Read the recorded cross-encoder score for each pair.
4
Sort those candidates again and compare the source evidence.