@swissspidy/belay-core

Answer AI tasks with a small on-device model, and call a cloud model only for the inputs it is unsure about. @swissspidy/belay-core has the task, the cascade, confidence handling, calibration files, cloud adapters and savings telemetry. It has no runtime dependencies.

Pair it with @swissspidy/belay-web for the on-device runners (Chrome's Classifier and Prompt APIs), and calibrate the threshold on your own data with @swissspidy/belay-calibrate.

Status: experimental (0.x). The API may change between minor versions.

Install

npm install @swissspidy/belay-core @swissspidy/belay-web

Usage

import { task, fetchAdapter } from '@swissspidy/belay-core';
import { classifierApi } from '@swissspidy/belay-web';

const triage = task({
  name: 'ticket-triage',
  schema: { type: 'categorical', options: ['bug', 'billing', 'feature', 'other'] as const },
  local: classifierApi(),
  cloud: fetchAdapter({ url: '/api/belay' }), // your backend calls the LLM
  threshold: 0.8, // or: calibration: '/belay.calibration.json'
});

const result = await triage.run(ticketText, { signal });
// { value: 'billing', confidence: 0.93, source: 'local', escalationReason: null, ... }

The local answer is accepted if and only if confidence >= threshold. Otherwise the task escalates to the cloud, subject to your privacy policy (redact, consent, or never).

Documentation

See the project README for schemas, confidence, privacy, cloud adapters, calibration and telemetry, and the ADRs for the design decisions.

License

Apache-2.0