AI WebGPU Lab Experiment

Embeddings Throughput Harness

`exp-embeddings-browser-throughput` now exposes a deterministic browser-only throughput harness. It embeds a fixed document fixture, records cold and warm runs, and measures query recall over the same cached index.

This is the first repo-specific baseline surface. Replace the synthetic vectorizer with a real runtime/model once the integration path is fixed.

Run Controls

Run cold first to build and persist the document index, then run warm to measure reuse against the same fixture.

Harness Scope

  • Fixed fixture with reproducible document order and query set.
  • Deterministic vectorizer so cold/warm deltas are attributable to caching, not model randomness.
  • Recall@10 is measured against expected source documents for the bundled query set.

Metrics

Environment

Activity Log

    Schema-Aligned Result Draft

    {
      "status": "pending"
    }