mirror of
https://github.com/ruvnet/RuVector.git
synced 2026-11-03 12:17:27 +00:00
cc0ff4b31a
sift1m-bench measures ruvector-core on TEXMEX fvecs/ivecs files, and the sota-bench SOTA check divides by a fixed hnswlib QPS constant rather than a measured one. This adds benchmarks/vector-search/hnswlib_compare.py: - `gen` writes a seeded Gaussian dataset with exact L2 top-100 ground truth in the file layout sift1m-bench reads. - `run` runs hnswlib and a numpy exact search on those files (or real SIFT1M) with the same M, ef_construction, ef sweep and k, and prints the same columns as sift1m-bench. hnswlib builds single-threaded by default, as sift1m-bench does. Each ef also reports a batched single-threaded QPS, which leaves out the per-call Python overhead. Replaces the harness from #293, which ran hnswlib and ruvector-core on different random datasets. Co-authored-by: RuFlo <ruv@ruv.net> |
||
|---|---|---|
| .. | ||
| docs | ||
| graph | ||
| src | ||
| vector-search | ||
| .dockerignore | ||
| .gitignore | ||
| Dockerfile | ||
| package.json | ||
| setup.sh | ||
| visualization-dashboard.html | ||