FIG. 05 // HSPANN.RECALL_COST FINEWEB · D=768 · N=1M→100M · Q=1,000
Recall against query cost, one curve per index size
Larger collections need a shrinking fraction of their codes scanned. Above 10M the curves draw together: search still gets more expensive, but far less than the collection grows.
recall@
Latency comes from quiet reruns on an otherwise idle machine, measured at 1M, 10M, 30M and 100M only.
One curve per size: the best recall achieved at or below each cost across the tested beam and rerank settings. Warm pass · one query at a time · 28 cores · r6i.32xlarge · exact answers computed for each size. Hover for a configuration.