decidekit docs

Start with the README on GitHub for a five-line example and the headline numbers.

Guides

• Training a head: collecting examples, how many per label, label names, retraining, shipping heads, the CLI and choosing a model.
• Confidence and thresholds: what calibrated means, how the temperature is fitted, picking minConfidence on your own data and "none of these".
• In the browser: the worker, download progress, cross-origin isolation, first-load size and keeping the encoder out of the page bundle.
• Falling back to an LLM: withFallback(), learning from the fallback, measured savings.
• React: useDecider() and useDecision().
• Other languages: Decide Multilingual, measured on Japanese, Chinese, German and Spanish.
• Coming from Jev: TypeSafe's client shape on a local decider, what differs and using Jev as the fallback.
• Bundlers: Vite, webpack and Next.js config.
• switch: Decide as a decision model in the switch harness, with trained heads and a teacher, and WikiRace results.
• Use cases: good fits, poor fits and example code.

Reference

• API reference: every export of decidekit and decidekit/browser, and the CLI.
• How it works: the encoder, the head and each training step, with the numbers behind them.
• Moving parts: every piece on one page (encoder, scoring, heads, calibration, fallback, worker, isolation, bundlers, caching, head files), where it lives and the numbers behind it.
• FAQ.
• Ready-made heads: BANKING77 and CLINC150 heads to try before collecting examples.

Write-ups

• Introducing decidekit: what it is and the headline numbers.
• When the LLM is wrong: how withFallback learns from an LLM that makes mistakes.