question-kitThe standard library for System One models

Built by measurement, not guesswork: methods developed against thousands of labeled cases, so you don't have to run the tests yourself. For Jev, from TypeSafe.

Get started

Build a service agent

GitHub

A service agent you configure

Describe a service as data (what customers come for, what to collect, the tools, the steps) and it runs the conversation turn by turn. Code runs the procedure; Jev reads the customer. No language model writes anything, and every reading passes a confidence gate first.

question-kit/service-agent

An order taker

One message in, a structured order out, for any menu you describe, plus whether to trust it. Code reads what the menu knows; Jev tags the rest and checks the result.

question-kit/order

The building blocks

Nineteen ready-made methods that write their own questions (classify, pick one of many, extract a value or a date, check a record, search, rank, screen and more), each returning a typed result with a confidence.

question-kit

WARNING

Alpha. question-kit is new, and its APIs will change between 0.x releases. Please try it and tell us what breaks or feels awkward: open an issue. We're not taking pull requests yet; if you'd like to contribute, open an issue first so we can talk it over.

What it's for

Jev doesn't write text. You send it some state and a batch of closed questions ("which of these?", "yes or no?") and it returns a probability for every option, fast and cheaply. That makes it a good fit for the decisions inside an application: what a customer wants, which line of an order they mean, whether a form is filled in right.

The hard part is writing the questions well and fitting code around the answers. question-kit does that part:

import { classify, run } from "question-kit";

const intent = await run(
  client,
  classify("where is my package? it was due Friday", {
    shipping: "Where an order is, or when it will arrive",
    refund: "Money back for an order",
    account: "Logging in, passwords, account details",
  }),
);

intent.value;      // "shipping"
intent.confidence; // 0.97

The methods follow TypeSafe's cookbooks, which we recommend reading: each is a recipe for getting one kind of answer out of Jev, and question-kit turns the recipes into functions you call with a little structured state.

For now question-kit works with Jev only. We haven't measured other decision models yet, and want to before supporting them.

Where to go next

Getting started: install, connect Jev, ask a first question.
Tutorials, start to finish:
- Sort a support inbox: five core methods composed into one request per message.
- Build a service agent: the service description, written in the order the needs come up.
Building an agent: the reference for the service description, one turn, and the loop. The pizza shop shows all of it at a counter.
How the agent decides: gates, confidences, and what happens when Jev isn't sure.
Python (preview): the core plumbing and a first set of methods, on PyPI as question-kit, and what's supported in each language.

TIP

Building with a coding agent? The package ships a skill for it: skills/question-kit/SKILL.md in node_modules/question-kit/ after an install, or npx skills add Engid/question-kit. It tells an agent how to think in Jev's questions, which method to reach for, and the traps.