Benchmarks
Benchmark Suite
Comprehensive kNN backend and metric comparison for chemical and embedding spaces. All runs share hardware, profiles, and evaluation protocol; toggle between backend comparison and USearch metric tabs below.
USearch (HNSW)
tmap2 LSH (Numba)
tmap1 LSH (C++)
Dataset:
Profile:
Recall@20 vs Scale
enamine / medium
Total Pipeline Time (s)
enamine / medium - log
Pre-layout / kNN Time (s)
enamine / medium - log
Peak RSS (MB)
enamine / medium
Recall vs Total Time
enamine - top-left = best
MST Weight / Node vs Recall
enamine - lower = better tree
Tree Topology: Hop Analysis
1-NN Hops vs Scale
enamine / medium - ideal = 1.0
Mean Hops (r1–r20)
enamine / medium
Mean Hops vs Recall
enamine
Hops-by-rank size:
Hops by rank (1–20) @ 500k
enamine / medium - ideal = 1.0
Summary size:
Summary @ 1M (enamine / medium)
| Backend | Recall@20 | Pre-layout (s) | Total (s) | Peak RSS (MB) | MST wt / node | Components | Mean hops |
|---|---|---|---|---|---|---|---|
| USearch (HNSW) | 96.4% | 45.48 | 98.45 | 8287 | 0.3772 | 4 | 71.86 |
| tmap2 LSH (Numba) | 59.4% | 7.47 | 79.94 | 9156 | 0.3857 | 1 | 69.42 |
| tmap1 LSH (C++) | 17.8% | 62.13 | 213.23 | 8487 | 0.4650 | 1 | 104.72 |