Local capacity baseline

This page records a repeatable local signal for capacity planning. It is not a production SLO: timings depend on hardware, thread settings, dependency versions, graph shape, model configuration, and query mix.

Environment

The baseline was captured on 2026-10-15 at commit 02e6917 using the local .venv-audit environment:

Results

Command:

source .venv-audit/bin/activate
pytest -q benchmarks/forward.py --benchmark-only --benchmark-min-rounds=3 --no-cov

The forward benchmark uses a 32-query batch and visual=16, text=8:

The model-backed serving smoke also passed locally:

make serve-model-smoke

It measured three synthetic recommendation requests at p50 1.0 ms, p95 1.2 ms, and p99 1.2 ms under the default 30-second p95/p99 smoke budgets. This measures the small bundled smoke model, not a representative production catalog or hardware deployment.

Interpretation and follow-up

These results establish a reproducible comparison point, not an acceptance threshold. Before production use, repeat the same measurements on the target CPU/GPU, catalog size, modality dimensions, concurrency, and request mix; then set explicit p95/p99 startup, retrieval, completion, recommendation, and memory budgets in the release gate. The current repository does not yet have that deployment-specific calibration.