themisim.run_benchmark

themisim.run_benchmark(artifacts, *, k=10, nprobe=64, prefilter=500, n_latency=50, n_cold=10, n_recall=50, brute_force=True, chunk_rows=1000000, sweep=True, sweep_nprobe=(1, 8, 32, 64, 256), sweep_prefilter=(100, 500, 1000, 2000), operating_points=((64, 500), (256, 1000)), diversify_settings=(0, 30), ground_truth_path=None, seed=0, on_progress=None)

Full report: environment, hyperparameters, latency, throughput, recall.

The brute-force pass is the expensive part — one sequential sweep of the whole memmap — so it runs last, after every measurement that would be distorted by saturating the same device.

Parameters:

• artifacts (str | Path)

• k (int)

• nprobe (int)

• prefilter (int)

• n_latency (int)

• n_cold (int)

• n_recall (int)

• brute_force (bool)

• chunk_rows (int)

• sweep (bool)

• sweep_nprobe (Sequence[int])

• sweep_prefilter (Sequence[int])

• operating_points (Sequence[Tuple[int, int]])

• diversify_settings (Sequence[int])

• ground_truth_path (str | Path | None)

• seed (int)

• on_progress (Callable[[str, object], None] | None)

Return type:

dict