Probe partitions without confusing them with answers

For a simple Euclidean IVF design, training learns nlist centroids from representative vectors. Adding a document assigns it to a centroid and stores its ID plus a representation inside that centroid’s inverted list. An IVFFlat list retains the original vector payload for exact distance computation within the lists that are scanned. Training quality matters because an unrepresentative sample can produce imbalanced or poorly placed cells, creating expensive hot lists and systematic routing misses.

A query is compared with the coarse centroids, and the selected nprobe lists supply candidates. Real implementations may use an approximate coarse quantizer or enforce extra scan limits; those create additional approximation boundaries. Separate these choices in your experiment. If a coarse quantizer misses a relevant list, exact scoring inside another list cannot recover the absent document. Likewise, reranking only the surviving candidates repairs distance estimates but not routing omissions.

The number of scanned entries is the sum of the selected list lengths when every selected entry is inspected once. A balanced-list estimate replaces each length by N/nlist, but that approximation can badly underestimate a heavy cluster. Record the distribution of selected-list populations per query, not only the configured nprobe. Two query classes can share the same setting while producing very different memory traffic and tail latency because their routes encounter different regions.

In the deliberately simple case of IVFFlat with every list selected, no scan truncation, the same eligible snapshot, and exact distances, every stored vector is evaluated. That recovers exhaustive ranking subject to the stated floating-point and tie contract. It also loses the candidate-pruning advantage and can be slower than an ordinary Flat scan. Searching all IVFPQ lists is still not exact original-vector search because compressed distance estimates remain; the representation and routing approximation must be diagnosed separately.

Build exact ground truth once for a frozen corpus and query set, then sweep nprobe while holding the trained centroids and stored representation fixed. Track recall, latency percentiles, and candidates inspected. This isolates the effect of visiting more lists from the effect of retraining or changing the data. Include queries near known cell boundaries and rare document groups: an average over the dominant cluster can conceal a severe loss for another population.

Training and evaluation should use disjoint query roles when parameters are selected. After choosing a setting on a validation set, assess it on held-out queries rather than reporting the best point selected on the final test set. When corpus drift changes list balance or encoder geometry, retune and potentially rebuild. Adding more vectors to a trained index is not equivalent to adapting its coarse centroids; their fixed positions may become a worse description of the new population.