npm run bench in the engine repository measures how accurately the parser recovers fields from resumes, and how long a parse takes. The corpora are invented resumes. There is no third-party or real-world benchmark yet.

What it measures

Field accuracy: for each labelled resume, whether the parser recovered each field correctly. Two corpora:

• Hand-written fixtures. The resumes in the repository's test fixtures, read as text. This is a regression check, since the engine was built against them.
• Generated resumes. 30 seeded invented resumes with varied separators, date forms, city placement, headings and section order. Each is rendered four ways: as text, as a PDF, as a two-column PDF and as a DOCX. They are labelled from their inputs and never from what the engine reads, and they are not tuned to the engine.

bench/baseline.json lists every field each corpus misses today. The run fails on a miss that is not listed, so one field read better cannot hide another read worse.

Results on 0.11.0

Timing over 142 parses, after one warm-up: median 4.3 ms, p95 31.2 ms, measured on 0.10.0 (0.11.0 changes no reading or scoring code). These were taken on a developer machine, and timing varies with the machine and the load.

Bundle sizes

Minified and gzipped, from npm run size, on 0.11.0. Importing only part of a subpath bundles less: withLocales alone is 48.7 KB.

/node runs on Node only. The other subpaths run in browsers and at the edge.

Run it yourself

git clone https://github.com/VeriWorkly/ats-engine.git
cd ats-engine
npm ci
npm run build
npm run bench

You need Node 22.12 or later. npm run bench -- --verbose lists the fields missed in the generated corpus. Run npm run size for the bundle table.