← All questions
Retrieval & EmbeddingsVery Hard· IVF-PQ· Latency· Performance Tuning

How do you adjudicate a proposal to increase nlist to r...

Increasing nlist from 4,096 to 65,536 significantly reduces the number of additions per query, which helps with tail latency. However, it also reduces the search coverage. To maintain recall, you must increase the number of probes (w). This creates a trade-off: the scan time decreases, but the residual table build time may increase significantly. The adjudication depends on whether the recall can be recovered by re-tuning w; if not, the original nlist configuration was likely more appropriate for the data distribution.

Code playground (Python)