@weaveintel/testing

Fakes and an eval runner for testing weaveIntel code without real models, vector stores, or network calls.

Why it exists

Testing code that calls a language model is like rehearsing a play with a stand-in actor: you don't want the real, expensive, unpredictable star for every rehearsal — you want a reliable stand-in who says exactly the line you scripted, every time. testing gives you those stand-ins: a fake model that returns what you told it to, a fake vector store, a fake transport, and a whole fake runtime — so your tests are fast, deterministic, and free of network flakiness. When you do want to grade quality against a rubric, the eval runner is there too.

When to reach for it

Reach for it in unit and integration tests, or anywhere you want deterministic behavior instead of live model output — CI, local dev, reproducing a bug. Use the ./evals subpath when you're scoring outputs against a rubric rather than just stubbing a dependency. For hand-built inline mocks aimed at DX scaffolding rather than test suites, see @weaveintel/devtools.

How to use it

import { weaveFakeModel, FakeRuntime } from '@weaveintel/testing';

const model = weaveFakeModel({ responses: ['Hello from the stand-in.'] });
const runtime = new FakeRuntime();

const res = await model.generate({
  messages: [{ role: 'user', content: 'hi' }],
});

console.log(res.content); // "Hello from the stand-in."

What's in the box

• Fakes — weaveFakeModel, weaveFakeEmbedding, weaveFakeVectorStore, weaveFakeTransport (fake MCP server transport), weaveFakeContainerRuntime.
• Fake runtime — FakeRuntime, with FakeModelOptions / FakeRuntimeOptions to script behavior.
• Container types — ContainerRuntime, ContainerRunResult.

Subpath entry point: @weaveintel/testing/evals — the eval runner and rubric scoring.

License

MIT.