SereneDB vs Qdrant — vector search benchmark

Head-to-head ANN comparison on the Yandex Text-to-Image dataset: SereneDB's IVF index (optionally quantized) against Qdrant's HNSW index, same base vectors, queries, and ground truth.

Loading run metadata…

How to read these charts

Different algorithms, different knobs. SereneDB = IVF (nprobe), Qdrant = HNSW (hnsw_ef) — not the same unit, so compare the recall-vs-QPS curve, not points at equal effort.
"Best per band" = highest QPS at that recall@10 across every config measured, ties broken by lower p50. Bands are 0.05-wide up to 0.95 recall, then 0.01-wide above it.
QPS is closed-loop at 32 concurrent clients, same protocol for both engines.
Build-time/index-size panels are per-engine, not head-to-head (different structures and knobs). SereneDB is pinned at nlist_factor=2 (RaBitQ only ran at nlist_factor=1, so it's absent here). The build-time bar stacks CREATE INDEX under its VACUUM (COMPACT_TABLE) merge from one build; index size shows compact vs. no-compact and includes the base-table columnstore copy (Qdrant's is index-only).

Recall vs. QPS

For each recall@10 band, the best achievable query throughput at 32 concurrent clients. Y-axis is log-scaled — at high recall, QPS drops by an order of magnitude for both engines.

Show data table

Build time

Time to a queryable index, shown separately per engine (see caveat above on why these aren't overlaid).

SereneDB — by quantizer

CREATE INDEX time, with the final VACUUM (COMPACT_TABLE) merge stacked on top. nlist_factor=2 (6,325 cells).

Show data table

Qdrant — by (m, ef_construct)

HNSW graph construction time; full precision vs. scalar quantization.

Show data table

Index size

On-disk footprint, shown separately per engine.

SereneDB index size

Includes the base-table columnstore copy of the vectors. nlist_factor=2 (6,325 cells).

Show data table

Qdrant index size

Index-only footprint; full precision vs. scalar quantization.

Show data table

Data generated by scripts/generate_pages_data.py from results/serenedb/summary.csv and results/qdrant/qdrant_tuning_all.csv. See the repo README for the full benchmark methodology.

Benchmark queries run against the Yandex Text-to-Image-1B dataset, licensed CC BY 4.0.