Benchmark suite comparing VelesDB against pgvector (HNSW).

• Microbench numbers are host-specific and must be reported with CPU, OS, Rust toolchain, and feature flags.
• Internal SIMD, sparse, and VelesQL cache measurements should be run with --features internal-bench.
• Do not claim superiority over FAISS, Qdrant, SimSIMD, or other systems unless the dataset, recall target, hardware, and methodology are matched.
• For the latest controlled-host remediation run, see benchmarks/results/2026-03-10-perf-remediation-report.md.

Legacy note: The table below was originally captured under v0.7.3 Docker benchmarks. For the latest controlled-host remediation run, see benchmarks/results/2026-03-10-perf-remediation-report.md.

Legacy comparison from v0.7.3 Docker benchmarks (2025). Hardware and methodology differ from current benchmarks. See DOCKER_BENCHMARK_RESULTS.md for details.

• SIMD AVX-512/AVX2 - 32-wide processing with FMA
• Adaptive HNSW params - HnswParams::for_dataset_size() for optimal recall
• Parallel search - Batch parallel with prefetching
• Quantization - SQ8 (4x) and Binary (32x) compression

docker-compose up -d --build  # Start both servers
python benchmark_docker.py --vectors 5000 --clusters 25

python benchmark_recall.py --vectors 10000

# 1. Start both servers (Docker required)
docker-compose up -d --build

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run fair Docker benchmark
python benchmark_docker.py --vectors 5000 --clusters 25

# Both scripts support:
--vectors 5000     # Dataset size
--dim 768          # Vector dimension  
--queries 100      # Number of queries
--clusters 25      # Data clusters (realistic)

# Docker benchmark only:
--velesdb-url http://localhost:8080

Both databases are measured with total time including index construction:

• VelesDB: Insert + inline HNSW indexing
• pgvector: Raw INSERT + separate CREATE INDEX time

This ensures an apples-to-apples comparison of the complete ingestion pipeline.

Use these commands for host-local remediation runs:

cargo bench -p velesdb-core --features internal-bench --bench simd_benchmark -- 768 --noplot
cargo bench -p velesdb-core --features internal-bench --bench sparse_benchmark -- sparse_insert --noplot
cargo bench -p velesdb-core --features internal-bench --bench velesql_benchmark -- velesql_cache --noplot

Report the exact host and toolchain alongside the measured values.

VelesDB uses adaptive parameters based on dataset size:

// Automatic parameter selection
let params = HnswParams::for_dataset_size(768, 100_000);
// Or for 1M scale
let params = HnswParams::million_scale(768);

MIT License