How do you choose between doubling the sliding window s...

Search’s proposal to double the window size often fails to meet strict latency targets because the reduction in calls does not offset the document slots processed. Retrieval’s proposal to cut depth is more effective for latency but directly impacts recall. The decision should be based on existing data regarding recall@50 versus recall@100. If recall@100 is significantly better, consider cascading a cross-encoder from 100 to 20 to maintain performance within the budget.