v0.3 Vector Performance

v0.3 ships a vector benchmark harness. It measures HNSW raw, HNSW with product quantization, IVF-flat at different probe counts, and brute-force exact search as the recall reference.

Run it from the repository:

cd benchmarks/vector-compare
cargo run --release

Default dataset:

Setting	Value
Vectors	50000
Dimensions	96
Queries	200
k	10

Current Snapshot

Latest local run:

Backend	Build ms	Recall@10	Mean us	p50 us	p95 us	p99 us
AionDB HNSW raw	36571	0.996	9423	8790	15674	17233
AionDB HNSW PQ	74742	0.994	13072	12355	17659	18471
AionDB IVF-flat nprobe=8	418	0.466	809	827	1223	1766
AionDB IVF-flat nprobe=32	416	0.863	2572	2474	3603	3977
Brute-force exact	0	1.000	9930	9520	14293	18991

What The Numbers Show

HNSW raw is the high-recall path: 0.996 recall@10 on the default run.

HNSW PQ keeps recall high at 0.994 recall@10 while exercising the product-quantized search path and exact rescoring.

IVF-flat is the fast-build path: about 416-418 ms build time on this dataset. With nprobe=32, it reaches 0.863 recall@10 at about 2.57 ms mean query latency.

Brute-force exact remains in the benchmark as the ground truth. That keeps every approximate result tied to a measurable recall target instead of a standalone latency number.

v0.3 surface

Vector performance is a normal evaluation surface in v0.3. SQL users create vector indexes through pgvector-style DDL. RAG workloads combine nearest-neighbor search with Qdrant-style metadata filters. Application data stays in one catalog with relational, graph, and vector access paths. Benchmark output includes build time, recall, and latency in one table.

The default run is self-contained. Provision external pgvector and Qdrant targets separately for service-to-service comparisons.

Raw Output

vector-compare  (n=50000  d=96  queries=200  k=10)

backend                        build_ms   recall@k    mean_us     p50_us     p95_us     p99_us
----------------------------------------------------------------------------------------------
aiondb hnsw (raw)                 36571      0.996       9423       8790      15674      17233
aiondb hnsw (pq)                  74742      0.994      13072      12355      17659      18471
aiondb ivf-flat (nlist=64,nprobe=8)        418      0.466        809        827       1223       1766
aiondb ivf-flat (nlist=64,nprobe=32)        416      0.863       2572       2474       3603       3977
brute-force (exact)                   0      1.000       9930       9520      14293      18991