Benchmarks

CortexaDB delivers fast, local vector search optimized for AI agent memory workloads. Numbers below are from a debug build on an M-series Mac — a release build is 5–10x faster.

Performance Overview

Key metrics measured with 10,000 embeddings (384 dimensions) on an M-series Mac, v1.0.0 debug build.

1.03ms p50 Search

HNSW search on 10,000 vectors. Release build achieves ~0.3ms p50.

952+ QPS

HNSW throughput (debug build). Release build exceeds 3,000 QPS.

95% Recall

Approximate search maintains high accuracy relative to brute-force.

Retrieval Benchmarks

Measured: 10,000 embeddings × 384 dimensions, 1,000 queries, 100 warmup, top-10.

Mode	p50	p95	p99	Throughput	Recall	Disk
HNSW	1.03ms	1.18ms	1.29ms	952 QPS	95%	47 MB
Exact	16.38ms	22.69ms	35.77ms	56 QPS	100%	31 MB

[!NOTE] These numbers are from a debug build (maturin develop). With a release build (maturin develop --release), HNSW achieves ~0.3ms p50 and 3,000+ QPS — consistent with the ingestion benchmarks below.

Ingestion

Operation	Time
Bulk Ingest (1,000 chunks)	0.12s
Single Memory Add	1ms
HNSW Index Build (10,000 vectors)	~286s (debug) / ~140s (release)

Methodology

Dataset: 10,000 random embeddings × 384 dimensions.
Environment: M-series Mac. Debug build via maturin develop.
Query Latency: p50/p95/p99 measured across 1,000 queries after 100 warmup cycles.
Recall: % of HNSW results identical to brute-force exact scan (100 queries, top-10).

Reproducing Results

Build the release extension for best performance:

cd crates/cortexadb-py
maturin develop --release
cd ../..
pip install numpy psutil

Run the automated benchmark suite:

# Generate 10k test vectors
python3 benchmark/generate_embeddings.py --count 10000 --dimensions 384

# Benchmark HNSW performance
python3 benchmark/run_benchmark.py --index-mode hnsw

# Benchmark Exact performance
python3 benchmark/run_benchmark.py --index-mode exact

Decision Guide

Metric	Use Exact	Use HNSW
Dataset Size	< 10,000 entries	> 10,000 entries
Recall Needed	100% (Strict)	95-99% (Semantic)
Latency Target	< 20ms (debug) / < 2ms (release)	< 5ms (debug) / < 1ms (release)
Resource Profile	Minimum Memory	High Performance

[!TIP] For datasets between 1k and 10k, Exact mode is often a good choice due to zero index-building overhead and 100% recall. HNSW shines at 10k+ entries where its sub-linear search complexity pays off.

Next Steps

Indexing - HNSW configuration and tuning
Configuration - All database options