Hierarchical SPANN — cost at a fixed recall

FIG. 06 // HSPANN.COST_AT_FIXED_RECALL FINEWEB · D=768 · N=1M→100M · Q=1,000

Cost to reach a fixed recall, by index size

Ten times more data needs far less than ten times the scan. Latency grows faster than the scan, because routing a larger tree adds work of its own.

recall@[

Both axes are logarithmic. Legend powers are fitted from 10M to 100M; shaded bands are 95% intervals over queries. The beam grid steps in factors of two, so the powers are approximate. Scans: all checkpoint sizes. Latency: median on an otherwise idle r6i.32xlarge at 1M, 10M, 30M and 100M, one query at a time. Hover for values.