Questions

A question declares what the application wants to know and which answers are allowed. Providers never receive free-form prompts from application code, only these declarations. That is what lets the same question be answered by a regular expression, a 200M parameter encoder or a frontier LLM without any change to the code that reads the answer.

The four kinds follow the System One wire format that TypeSafe introduced for Jev and that Kev, OpenJev and Laya also implement, plus Extract.

Choice

Exactly one option out of a fixed set.

from thinkless import Choice

INTENT = Choice(
    "What does the customer want?",
    name="intent",
    options={
        "refund_duplicate_charge": "was billed twice for one purchase and wants the extra charge back",
        "order_status": "asks where an order is, when it arrives, or about tracking",
        "other": "anything else",
    },
)

The decision value is the chosen label. decision.probabilities holds the full distribution when the provider has one.

Options can also be a plain list of labels. Descriptions are worth writing: every provider, from GLiNER to an LLM, uses them to tell options apart. Write them the way you would brief a new teammate, and describe the boundary between similar options explicitly.

Score

A position on an ordered scale, lowest first.

from thinkless import Score

URGENCY = Score("How urgent is the request?", name="urgency", levels=("low", "medium", "high"))

The value is the expected level index as a float (for example 1.7), and decision.level is the single most likely level.

YesNo

A binary question. The value is a bool.

from thinkless import YesNo

WANTS_HUMAN = YesNo(
    "Does the customer explicitly ask to talk to a human instead of an automated reply?",
    name="wants_human",
    yes_means="asks for a person, an agent, a callback or a manager",
    no_means="anything else, including frustration without a request for a person",
)

yes_means and no_means are optional. They help most on questions whose boundary is subtle.

Extract

Named fields pulled out of the input.

from thinkless import Extract

ORDER = Extract(
    name="order",
    fields={"order_id": "the order number the customer refers to, four or five digits"},
    required=["order_id"],
)

The value is a dict of field values, with None for fields that were not found. Fields listed in required must be present for the answer to count as confident; a missing optional field is simply None. Extraction replaces one of the most common LLM calls in agents: turning a message into tool arguments.

Options shared by every question

Option	Meaning
name	Stable identifier used in traces, per-question thresholds and rules. Derived from the instructions when omitted, but set it explicitly for anything you depend on.
threshold	Minimum normalized confidence to accept an answer. Overrides the engine default.
providers	Allowlist of provider names that may answer. Use it to keep a question away from a model that cannot answer it well.

Questions are immutable. Define them once, at module level, and reuse them.

Asking several questions at once

decisions = engine.decide_many(ticket, [INTENT, URGENCY, WANTS_HUMAN, ORDER])

The questions travel through the cascade together. Each provider receives one call with every open question it supports, which is how System One models are designed to be used (one forward pass answers them all) and what keeps an LLM-only baseline fair: one prompt answers the whole batch.

Writing questions that small models answer well

• Ask about one thing. "Is the customer angry and asking for a refund?" is two questions.
• Name the population in the instructions: "the customer", "the message", "the help article".
• Prefer a Choice with an explicit other option over a set of yes/no questions that could all be true.
• Keep labels short and literal. The label is what your code switches on; the description is where the nuance goes.
• Measure, do not guess. Calibration tells you whether a model can answer a question and at what threshold.