📈Pro workspace · Vector search

HNSW Tuning

Tune the graph that powers approximate nearest-neighbour search. See how M, ef_construction, and ef_search trade recall, latency, memory, and build time.

LevelAdvanced

Time20 min

FocusRecall vs latency

Pro tuning workspace

Choose an HNSW operating point with evidence.

Pro members can sweep HNSW parameters, read the simulated Recall@10 and P95 curve, compare precision tiers, and practice the benchmark workflow used before a production index release.

Create a free account Read the tuning guide

Representative sweep

1M vectors · 768 dimensions

ef_search	Recall@10	P95
50	~92%	~25ms
100	~96%	~50ms
200	~98%	~100ms

These are teaching estimates, not live OpenSearch metrics. The Pro workspace makes the trade-offs visible; final tuning always uses exact ground truth on your own corpus.

Learn the HNSW theory behind the workspace

kNN and HNSW: How Vector Search Scales

Read article

HNSW Tuning: Find the Recall, Latency, and Cost Operating Point

Read article

Quantization: Cut Vector Memory by 4–40x

Read article