Documentation

Build with the Envariant SDK

Everything you need to add an interpretability and control layer to any model call. Access to live endpoints is granted off the waitlist.

1Installation

Add the SDK to your project with your package manager of choice. TypeScript and Python clients share the same concepts and API surface.

npm install @envariant/sdk
# or
pip install envariant

Create a client scoped to a model. Every trace, steer, and invariant call runs through this instance.

import { Envariant } from '@envariant/sdk'

const ev = new Envariant({ model: 'gpt-frontier' })

Trace a prompt to surface the concepts and internal features that drove the model's output.

const { features } = await ev.trace(prompt)

Nudge a named concept up or down at inference time without retraining or prompt hacking.

ev.steer('verbosity', -0.4)

Assert rules that must always hold in production. Violations are blocked and reported to your dashboard.

ev.assert('no_pii_leakage')