The p99 latency budget dropped significantly. How do yo...

The adjudication should be based on measured impact. Cutting candidates from 1,000 to 50 saves almost no time on retrieval but takes a 20x bite out of the reranker’s visibility. The optimal path is to keep both stages, reduce rerank depth to where the quality curve flattens, or move to late interaction if the budget is still exceeded. If first-stage recall at 50 is nearly identical to recall at 1,000, then the cut is free.