edgewise/test

Testing your app

edgewise/test makes mock models that answer instantly, so you can unit-test app logic without downloads.

import { generate, evaluate, choice } from 'edgewise';
import { mockModel } from 'edgewise/test';

const lm = mockModel({ verb: 'generate', respond: ({ messages }) => `echo: ${messages.at(-1)?.content}` });
const { text } = await generate({ model: lm, input: 'hi' }); // 'echo: hi'

const judge = mockModel({ verb: 'evaluate', respond: () => ({ topic: { billing: 0.8, other: 0.2 } }) });
const r = await evaluate({ model: judge, state: 'x', questions: { topic: choice({ billing: 'billing', other: 'other' }) } });
r.answers.topic.choice; // 'billing'

A mock is a real manifest, so it goes through the same resolution, Runs, streaming, schemas and tool loop as a real model. A generate mock can return tool calls:

const caller = mockModel({ verb: 'generate', respond: ({ messages }) =>
  messages.some((m) => m.role === 'tool') ? 'Done.' : { toolCalls: [{ name: 'setVolume', input: { level: 30 } }] } });

Mocks exist for generate, evaluate, embed, speak and forecast. An evaluate mock's respond receives { state, images, audio, sampleRate }, so you can test code that judges media.

Testing with real models#

Edgewise's own real-model tests are in test/models. Run them with bun run test:models. They download about 5 GB the first time.