A colleague wants to raise nprobe from 16 to 128 for be...

Increasing nprobe by 8x results in roughly 8x more data transfer, which typically adds 7.5–9.6 ms of latency. If this fits within the current 20 ms window, the change is safe. However, you should not hardcode this value; instead, gate it on live latency measurements and ensure that the recall improvement is empirically validated on an evaluation set.