Vector Database Benchmarks

This page records a local run of the cross-engine benchmark suite in benchmarks/chroma_lancedb_qdrant_lynsedb.

Benchmark results are environment-sensitive. Treat these numbers as one reproducible reference run, not as universal performance guarantees. RSS deltas, startup latency, and short resource-run ingest throughput are especially sensitive to allocator state, import order, and OS cache behavior.

Run Details

Item	Value
Date	2026-06-20
Command	python benchmarks/chroma_lancedb_qdrant_lynsedb/bench.py all --engines lynsedb-f32,chroma,lancedb,usearch --n 100000 --queries 100 --dim 128 --k 10 --batch-size 1000 --warmup 5 --jsonl benchmarks/chroma_lancedb_qdrant_lynsedb/results-2026-06-20-rerun.jsonl
Dataset	100,000 vectors, 128 dimensions, 100 queries, top-k 10
Python	3.12.2
Platform	macOS 26.5.1, arm64
LynseDB	0.7.1
ChromaDB	1.5.9
LanceDB	0.33.0
USEARCH	2.25.3
Raw results	benchmarks/chroma_lancedb_qdrant_lynsedb/results-2026-06-20-rerun.jsonl

The suite uses normalized synthetic vectors and exact NumPy inner-product ground truth for recall checks. This run uses LynseDB float32 storage with FLAT-IP, ChromaDB's persistent local HNSW collection, LanceDB's normalized vector table, and USEARCH's float32 HNSW index. Engines are ingested through their batch APIs.

Summary

In this 100k-row float32 run, LynseDB leads persisted batch ingest and exact vector, filtered, and hybrid-search latency among the database adapters while maintaining 1.000 recall@10. LanceDB uses the least disk. ChromaDB and USEARCH return approximate results in this configuration, so their latency must be read alongside recall.

100k Float32 Comparison

A run on 2026-06-20 compares LynseDB in float32 mode with ChromaDB, LanceDB, and USEARCH:

python benchmarks/chroma_lancedb_qdrant_lynsedb/bench.py all --engines lynsedb-f32,chroma,lancedb,usearch --n 100000 --queries 100 --dim 128 --k 10 --batch-size 1000 --warmup 5 --jsonl benchmarks/chroma_lancedb_qdrant_lynsedb/results-2026-06-20-rerun.jsonl

LynseDB float32 uses the same FLAT-IP index and exact recall target as the main embedded run. USEARCH 2.25.3 is included as a vector-only HNSW index with metric="ip", dtype="f32", and USEARCH_EXPANSION_SEARCH=128 default. It does not implement metadata filters or text/vector hybrid search in this suite, so those rows are skipped instead of emulated in Python.

Metric	LynseDB-f32	ChromaDB	LanceDB	USEARCH
Batch ingest vectors/s	73,399	2,108	68,123	10,578
Ingest time	1.36 s	47.45 s	1.47 s	9.45 s
Disk after ingest MB	69.13	162.42	55.76	63.03
Vector search mean ms	0.661	1.233	14.581	0.555
Vector search QPS	1,513	811	69	1,803
Vector search recall@10	1.0000	0.5180	1.0000	0.6000
Recall run mean ms	0.609	1.110	14.996	0.531
Recall run QPS	1,643	901	67	1,882
Recall run recall@10	1.0000	0.5180	1.0000	0.5970
Filtered search mean ms	0.178	37.354	16.692	n/a
Filtered search QPS	5,618	27	60	n/a
Filtered recall@10	1.0000	0.9990	1.0000	n/a
Hybrid search mean ms	4.809	n/a	17.810	n/a
Hybrid search QPS	208	n/a	56	n/a
Hybrid topic hit rate	0.6050	n/a	0.6000	n/a
Resource-run vectors/s	70,042	1,937	75,840	11,359
Startup mean ms	2.087	13.995	2.251	0.036

In this float32 validation, LynseDB leads persisted ingest and exact vector, filtered, and hybrid-search latency among the database adapters. USEARCH's HNSW search is slightly faster than LynseDB's exact flat search, but at much lower recall and without database-level metadata or hybrid retrieval features.

1M Row Scale Check

A larger single-machine check was also run on 2026-06-20 with 1,000,000 vectors, 128 dimensions, 100 queries, top-k 10, and batch insert APIs:

python benchmarks/chroma_lancedb_qdrant_lynsedb/bench.py all --engines lynsedb-f32,lancedb --n 1000000 --dim 128 --queries 100 --k 10 --batch-size 5000 --warmup 5 --jsonl benchmarks/chroma_lancedb_qdrant_lynsedb/results-2026-06-20-1m-rerun.jsonl

ChromaDB is not included in this 1M summary because the recorded run failed under the selected batch-size settings. LanceDB completed the scale check, so the table compares LynseDB and LanceDB directly.

Metric	LynseDB	LanceDB
Batch ingest vectors/s	49,954	85,057
Ingest time	20.02 s	11.76 s
Disk after ingest MB	694.32	547.69
Vector search mean ms	6.013	109.009
Vector search QPS	166.31	9.17
Vector search recall@10	1.0000	1.0000
Recall run mean ms	5.573	148.741
Recall run QPS	179.42	6.72
Filtered search mean ms	2.160	148.455
Filtered search QPS	463.03	6.74
Filtered recall@10	1.0000	1.0000

This run shows LynseDB handling 1 million 128-dimensional float32 vectors on a single machine: ingest completes in about 20.0 seconds, persisted storage is about 694 MB, exact vector search stays around 6.0 ms mean latency, and metadata-filtered search stays around 2.2 ms mean latency. LanceDB ingests faster and uses less disk, while LynseDB has lower exact and filtered-search latency.

Persisted Vector + Metadata/Text-Index Ingest

This ingest metric includes committed vector writes, metadata persistence, and the text index needed by the hybrid-search benchmark. It is not a pure vector-only append microbenchmark.

Engine	Vectors/s	Disk MB	RSS delta MB
lynsedb-f32	73,399	69.13	399.83
chroma	2,108	162.42	218.95
lancedb	68,123	55.76	149.02
usearch	10,578	63.03	70.38

Vector Search

Engine	Mean ms	P50 ms	P95 ms	QPS	Recall@10	Disk MB
lynsedb-f32	0.661	0.640	0.868	1,513	1.0000	69.13
chroma	1.233	1.197	1.525	811	0.5180	162.33
lancedb	14.581	14.281	16.845	69	1.0000	55.75
usearch	0.555	0.532	0.698	1,803	0.6000	63.03

Recall

Engine	Mean ms	P50 ms	P95 ms	QPS	Recall@10	Disk MB
lynsedb-f32	0.609	0.596	0.698	1,643	1.0000	69.13
chroma	1.110	1.068	1.360	901	0.5180	162.38
lancedb	14.996	14.770	17.297	67	1.0000	55.75
usearch	0.531	0.527	0.654	1,882	0.5970	63.03

Filtered Vector Search

Engine	Mean ms	P50 ms	P95 ms	QPS	Filtered recall@10	Disk MB
lynsedb-f32	0.178	0.168	0.257	5,618	1.0000	69.13
chroma	37.354	37.247	39.086	27	0.9990	162.35
lancedb	16.692	16.500	18.656	60	1.0000	55.75

Hybrid Search

Engine	Mean ms	P50 ms	P95 ms	QPS	Topic hit rate / note	Disk MB
lynsedb-f32	4.809	2.925	16.042	208	0.6050	69.13
lancedb	17.810	17.661	20.329	56	0.6000	55.77

topic_hit_rate is a synthetic sanity check for the benchmark's generated topics. It is not a relevance score for a production retrieval workload.

Resources

Engine	Resource-run vectors/s	Disk MB	RSS delta MB
lynsedb-f32	70,042	69.13	0.02
chroma	1,937	162.40	84.09
lancedb	75,840	55.75	5.52
usearch	11,359	63.03	67.61

RSS deltas are measured in-process and can vary with allocator state and import order. Negative deltas can occur when a later run reuses allocator state or the runtime releases memory during the measurement. Use disk size and repeated runs when comparing storage footprint.

Startup

Engine	Startup mean ms	Startup P50 ms	Startup P95 ms	QPS
lynsedb-f32	2.087	1.900	2.746	479
chroma	13.995	14.085	14.826	71
lancedb	2.251	2.201	2.706	444
usearch	0.036	0.023	0.073	27,868

LynseDB HTTP

Engine	Mean ms	P50 ms	P95 ms	QPS	Recall@10	Disk MB
lynsedb-http	2.439	2.163	4.812	410	0.9990	44.71

The HTTP row starts a local LynseDB service, ingests the same dataset remotely, and runs the same query workload through the client.