Benchmark reproducibility¶
- Benchmarks run via
pytest-benchmark; results stored underartifacts/test/benchmarks. - Use the same Python version and dependencies (see
pyproject.toml) to compare runs. - Do not compare benchmarks across machines unless normalized by CPU; instead, use the regression thresholds enforced in
tests/perf/test_retrieval_benchmark.py. - Always run with
PYTHONDONTWRITEBYTECODE=1to avoid filesystem noise.