bench/ef_search.sh results

What measure() reports for one hnsw index as hnsw.ef_search changes. Method: the script generates 20,000 vector(64) rows, uniform in [-1, 1], from setseed(0.42); builds an hnsw index (vector_cosine_ops, default m 16 and ef_construction 64); then, for each ef_search, runs measure('items_hnsw', 10, 100) three times (k = 10, 100 query vectors drawn at random from the table, so the three runs differ).

  • Date: 2026-11-12
  • Code: pg_recall_guard 0.2.9 (measure() is unchanged in 0.2.10); bench/ef_search.sh as committed in 142564a.
  • PostgreSQL 13.23, built from REL_13_STABLE 05f6c9ec2c4, CFLAGS=-O2, no assertions; pgvector 0.8.6.
  • Machine: AMD Ryzen AI 9 HX 370, Fedora, Linux 7.3.3. Recall does not depend on the machine; it depends on the data, the index build and the sample.
  • Command: PGHOST=<socket dir> PGPORT=<port> bash bench/ef_search.sh
hnsw.ef_search measured recall, 3 runs
1 0.0020, 0.0030, 0.0030
10 0.8460, 0.8470, 0.8250
40 (default) 0.9100, 0.9120, 0.9140
100 0.9540, 0.9490, 0.9520

These replace a table the README carried without a script behind it (PostgreSQL 19beta2, pgvector 0.8.6 and pgvectorscale 0.9.0, 38,352 vector(768) rows: 0.0000 / 0.9000 / 1.0000).