Same recall — up to 5.3× faster queries Median query latency on GloVe-100-angular (1,183,514 vectors, cosine) · lower is better pgContext pgvector 0 2 4 6 8 10 12 14 p50 latency (ms) 0.67 2.56 3.8× ef 64 recall 0.73 0.94 4.14 4.4× ef 128 recall 0.81 1.46 7.09 4.9× ef 256 recall 0.87 2.44 12.97 5.3× ef 512 recall 0.91 GloVe-100-angular, dataset ground-truth neighbors · both engines in one PostgreSQL 17 container, matched 8-way parallel build · Apple M4 Pro (NEON kernels) Reproduce: benchmarks/pgvector_comparison/run-ann-hdf5.sh · regenerate this chart: scripts/generate-benchmark-chart.py · 2026-07-20