question-kit for Python

The standard library for System One models. For now that means Jev, from TypeSafe.

Built by measurement, not guesswork: the core methods are developed against thousands of labeled cases, from public benchmarks and our own test sets, and tuned on what the results show.

Early release. This is the first piece of the Python port: the core plumbing and a first set of methods. The TypeScript package is the full kit (every core method, the order taker and the service agent), and the rest is coming to Python. APIs may change between 0.x releases; issues welcome.

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. The hard part is writing the questions well and fitting code around the answers. question-kit does that part: each method writes its own questions, following TypeSafe's cookbooks, and returns a typed result with a confidence.

Install

pip install "question-kit[typesafe]"

Python 3.10 or later. The typesafe extra installs TypeSafe's SDK (typesafe-sdk), which question_kit.typesafe wraps; leave it out if you bring your own client. Calls to Jev need TYPESAFE_API_KEY.

A first question

from question_kit import classify, run
from question_kit.typesafe import typesafe_jev

client = typesafe_jev()  # reads TYPESAFE_API_KEY

intent = 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", or "none" when nothing fits
intent.confidence  # 0 to 1, by TypeSafe's confidence formula

Several questions about the same text go in one request with run_all, the way TypeSafe recommends:

from question_kit import check, classify, extract_value, run_all

r = run_all(client, {
    "intent": classify(text, labels),
    "order": extract_value(text, kind="number", role="the order number"),
    "upset": check(text, "The customer is upset."),
})
r["order"].value   # "4410982": always text that's really in the message
r["upset"].value   # True or False, with r["upset"].confidence

What's in 0.1

The API is synchronous for now.

Compared with the TypeScript package

The two build byte-for-byte the same requests (a test checks this against requests built by the TypeScript package), so they share one answer cache: answers recorded by either replay in both.

Testing without Jev

from question_kit import fake_jev, run, check
from question_kit.cache import cached_jev

fake = fake_jev(lambda qid, question, state: "yes")       # every Noul says yes
run(fake, check("hi", "It's a greeting.")).value          # True

client = cached_jev(typesafe_jev(), ".cache/jev")          # record: ask once, keep the answer
replay = cached_jev(None, ".cache/jev", mode="replay")     # never calls the API; a miss raises

Links

• Docs: docs.questionkit.dev (the Python pages are under Python)
• Source: github.com/Engid/question-kit, in python/
• TypeSafe's docs: docs.typesafe.ai

MIT licensed. © 2026 Nick Gideo.