PerformanceBENCH-01 · SIFT1M · certification hardware

Bench
marks

A formal Python-native performance benchmark suite captures ingestion throughput and query latency under realistic single-threaded workloads. No network. No warm-up tricks. In-process, on bare metal.

95ingestion ops/secSDK · 10K · 128d

1.2msHNSW p50 · 10KRust Core · 128d

2.80xbest speedupSIFT1M Balanced L2

3,636QPS peakBalanced Cos 100K

§01

BENCH-01 · SDK Performance Baseline

10K records · 128 dimensions · Cosine · single-threaded · Python SDK (PyO3 boundary)

certified

Metric	p50	p99	Throughput
Ingestion (PUT)	13.174 ms	18.504 ms	74 ops/sec
Search (Lexical BM25)	N/D (outlier)	—	—
Search (Vector HNSW)	2.024 ms	4.403 ms	494 qps
Search (Hybrid Fusion)	3.114 ms	5.507 ms	321 qps

Query latency · p50 vs p99

lower is faster →

HNSW (Vector)p50 2.024ms · p99 4.403ms

Hybrid Fusionp50 3.114ms · p99 5.507ms

p50 p99scale 0–8ms

Hardware profile: 12-core CPU @ 3.5GHz, AVX2 enabled, Windows 11 / Ubuntu 22.04 LTS.

Source:docs/operations/BENCHMARKS.md §2 · reproduce: python benchmarks/vantadb_local_bench.py --size 10000 --dim 128 --queries 1000

Latency Explorer

Adjust the workload parameters and watch how VantaDB compares against a typical network-bound database. Numbers scale from the BENCH-01 baseline using an illustrative cost model.

Presets:

top_k5

Results returned per query

dimensions128d

Embedding vector width

dataset10K

Indexed vector count

p99 latency · lower is faster

scale 0–14.8ms

VantaDB · Hybrid2.9ms

◆

p50: 1.64msp99: 2.9ms

qps: 609

VantaDB · HNSW2.32ms

◆

p50: 1.07msp99: 2.32ms

qps: 937

Network DB (typical)14.76ms

p50: 6.59msp99: 14.76ms

qps: 152

throughput · queries/sec

Hybrid

609

HNSW

937

Network DB (typical)

152

Simulate benchmark

edge workload · 10K

Run a 3-second simulated benchmark of the current workload configuration. Results are extrapolated from the BENCH-01 baseline with ±4% jitter.

Note: Baseline from BENCH-01 (10K vectors, 128d, cosine). The cost model is illustrative — actual figures depend on hardware, AVX2, and ef_search configuration. Run python benchmarks/vantadb_local_bench.py for real numbers on your machine.

§02

SIFT1M · Phase 2 Competitive Benchmarks

Static prefetch · Euclidean sqrt elimination · pure SIMD cosine · O(M²) select_neighbors optimization

certified

Scale	Config	Metric	Before	Now	Speedup	p99	QPS
100K	Balanced Cos	Cosine	139.4s	63.7s	2.18x	441.2 µs	3,636
100K	High Recall Cos	Cosine	390.8s	182.2s	2.14x	1,231.8 µs	1,379
100K	Balanced L2	Euclidean	191.4s	68.4s	2.80x	671.4 µs	3,270
100K	High Recall L2	Euclidean	462.2s	194.5s	2.37x	1,183.6 µs	1,353
100K	High Recall L2 Mmap	Mmap Euclidean	411.2s	189.8s	2.16x	1,094.8 µs	1,438

Phase 2 construction speedups

SIFT1M · 100K vectors

Balanced Cos

2.18x

High Recall Cos

2.14x

Balanced L2

2.80x

High Recall L2

2.37x

High Recall L2 Mmap

2.16x

Optimizations: static prefetch · elimination of Euclidean square root in hot graph traversal · pure SIMD cosine similarity · O(M²) select_neighbors (caches references to eradicate HashMap queries during the diversity loop).

Certification hardware: AMD Ryzen 12-Core @ 3.5GHz, compiled with -C target-cpu=native.

Source:docs/operations/BENCHMARKS.md §5 · SIFT1M 100K · AMD Ryzen 12-Core @ 3.5GHz, -C target-cpu=native

§03

JS/WASM · vantadb-ts Bench

2K records · 384 dimensions · WASM in-memory in Node · seed 42 · median of 3 runs

certified

Operation	p50	p95	p99
Insert (putBatch)	667.61 ms	1,065.11 ms	1,453.48 ms
Search (Vector)	3.84 ms	10.97 ms	14.28 ms
Search (Hybrid)	184.04 ms	377.25 ms	447.18 ms

Hardware profile: Intel i5-1235U, Windows 11, Node.js v26.8.1. WASM engine is in-memory in Node (no OPFS); browser backends pending.

Source:docs/operations/BENCHMARKS.md §15 · 2026-09-07 · reproduce: cd vantadb-ts && npm run bench

§04

Competitive benchmark · embedded engines

glove-100-angular · 10K vectors · 100 queries · top_k=10. Measured run documented in docs/operations/BENCHMARKS.md §7 (2026-06-06); VantaDB runs through PyO3 over its mmap Rust core.

certified

Metric	VantaDB	LanceDB	ChromaDB	Pinecone	Weaviate
Ingest QPS	598.3	114,583	3,886	Managed	Managed
Index time (ms)	16,039.9	602.2	N/A (inc)	Managed	Managed
Query QPS	24.3	320.5	978.6	Managed	Managed
Latency p50 (ms)	39.74	2.653	0.941	Managed	Managed
Latency p99 (ms)	58.245	6.98	3.349	Managed	Managed
Recall@10	24.50%	13.90%	24.10%	Managed	Managed
Peak RSS (MB)	236.5	344.2	253.5	Managed	Managed
Delta RSS (MB)	91.7	97.2	39.1	Managed	Managed
Deployment	pip install · embedded	pip install · library	pip install · client/server	Fully managed SaaS	Self-host or cloud
Pricing model	Open source · $0	Open source	Open source	Usage / pods	Hosted by vendor
Durability	WAL + CRC32C	Arrow, query-versioned	In-memory HNSW (no WAL)	Managed SLA	WAL / hosted

Pinecone and Weaviate are managed/hosted services — the harness does not run them locally, so their measured cells read 'Managed'. We do not invent cross-vendor QPS/latency numbers here; all measured values come from BENCHMARKS.md §7 and are reproducible via the public harness.

Source:docs/blog/benchmarks_vs_lancedb_chroma.md

§05

Run the benchmark locally

Measure the baseline on your own hardware in three commands

certified

Install python bindings

pip install maturin
maturin develop --release

Execute the benchmark

python benchmarks/vantadb_local_bench.py \
  --size 10000 --dim 128 --queries 1000

Read the report

cat vanta_benchmark_report.json

Ready to build?

Install vantadb-py and run your first hybrid search in 5 minutes.

Benchmarks · Live Race

VantaDB vs the field.

Carrera visualizada de latencia p50 (HNSW · 10K). VantaDB 1.2ms certificado — competidores sin número local publicado (—). Press Start y mirá.

Head-to-head · live race

Standings

1VantaDB

2Qdrant

3Chroma

HNSW Vector Search · p50 Latency (ms)

VantaDB

1.2ms

Qdrant

—

Chroma

—

Weaviate

—

Pinecone

—

Recall@10 (mayor es mejor)

VantaDB

0.998

Chroma

—

Qdrant

—

Weaviate

—

Pinecone

—

VantaDB: BENCHMARKS §1 p50 1.2ms (10K) y Recall@10 0.998. Competidores: sin número medido publicado — la carrera es ilustrativa. Fuente: docs/operations/BENCHMARKS.md

Reproducible · JSON contract

VantaDB vs LanceDB & ChromaDB

glove-100-angular · 10,000 vectors · 100 queries · top-10 · cosine · single build

run 2026-06-06 · real

Engine	Ingest QPS	Index (ms)	Query QPS	p50 (ms)	p99 (ms)	Recall@10	Peak RSS (MB)	Δ RSS (MB)
VantaDB	598.3	16,039.9	24.3	39.740	58.245	24.50%	236.5	91.7
LanceDB	114,583.0	602.2	320.5	2.653	6.980	13.90%	344.2	97.2
ChromaDB	3,886.0	N/A (inc)	978.6	0.941	3.349	24.10%	253.5	39.1

Readings are real numbers from the certified run (2026-06-06) — glove-100-angular, 10,000 vectors, 100 queries, top-10. VantaDB runs through its PyO3 bindings over the mmap Rust core; LanceDB and ChromaDB use their native C/C++ wrappers. Regenerate: python benchmarks/competitive_bench.py --dataset glove-100-angular --size 10000 --queries 100

Source: docs/blog/benchmarks_vs_lancedb_chroma.md — repo-tracked JSON: web/src/lib/data/competitive-benchmark.json