Your p95 budget just dropped from 800 ms to 250 ms. You...

While 25 tokens fits the latency budget, it provides a fabricated span that is barely more ‘document-shaped’ than the original question, rendering the HyDE mechanism ineffective. The better approach is to either cache results if the hit rate is high, distill the generator into a query encoder, or abandon HyDE entirely to invest the serving budget into improving the underlying retrieval labels.