What rerank depth do you use, and why that number?

A strong answer provides an exchange rate: depth costs 0.195 ms per candidate for a BERT-base cross-encoder at 288 tokens, while recall gains vary based on whether you use a dense retriever or BM25. A weak answer relies on ‘as deep as latency allows’ or arbitrary numbers from papers, failing to account for the fact that past the knee of the curve, more candidates introduce distractors, which lowers quality while increasing costs.