Recall is low on an IVF-PQ index. What levers can you a...

To improve recall without increasing memory, start with residual encoding, which captures information PQ cannot model. Before implementing, check the correlation coefficient (ρ) from training to predict effectiveness. If recall is still insufficient, consider increasing the number of probes (nprobe) or the shortlist size, keeping in mind the latency impact. Increasing the number of sub-quantizers (m) is the final option, but it is not free in memory and should be used only when other methods are exhausted.