README.md

Benchmarks

Run from the repository root:

swift run -c release vectura-hnsw-benchmark

Environment options:

Variable	Default	Meaning
VECTURA_HNSW_BENCH_DOCS	2000	Number of documents/vectors
VECTURA_HNSW_BENCH_DIM	128	Vector dimension
VECTURA_HNSW_BENCH_QUERIES	25	Query count
VECTURA_HNSW_BENCH_TOPK	10	Requested result count
VECTURA_HNSW_BENCH_CANDIDATE_MULTIPLIER	8	HNSW prefilter multiplier
VECTURA_HNSW_BENCH_M	16	HNSW neighbor count
VECTURA_HNSW_BENCH_LEVEL0_MULTIPLIER	2	Ground-layer neighbor multiplier
VECTURA_HNSW_BENCH_LEVEL0_CAP	32	Ground-layer neighbor cap; use 0 for uncapped
VECTURA_HNSW_BENCH_EF_CONSTRUCTION	200	Insert-time search breadth
VECTURA_HNSW_BENCH_EF_SEARCH	128	Query-time search breadth
VECTURA_HNSW_BENCH_EXACT_THRESHOLD	10000	Document-count threshold for exact candidate fallback
VECTURA_HNSW_BENCH_BATCH_INSERTION_SEED	unset	Optional deterministic batch insertion shuffle seed

The benchmark reports:

• exact scan latency
• public VecturaHNSWKit candidate lookup latency
• HNSW-backed VecturaKit latency
• candidate recall@K before exact rescoring
• recall@1
• recall@K
• insert time
• snapshot write time
• cold open time
• snapshot size

The synthetic vectors are generated from stable seeds so repeated release runs use the same corpus and query set.

See RESULTS.md and OPTIMIZATION_RESULTS.md for local result snapshots.