Questions

A question describes one judgment. Four helpers create them: choice, multiLabel, score, and check. Each helper validates its input immediately and returns a plain data object tagged with its type.

import { check, choice, multiLabel, score } from '@anvia/core/decision'

Helper	Answer	Semantics
choice({ instructions, options })	choice, optional probabilities and confidence	Exactly one option is selected. Probabilities are exclusive and sum to one.
multiLabel({ instructions, options, threshold? })	labels, probabilities	Each label has an independent probability. Labels at or above the threshold are selected.
score({ instructions, rubric })	score, rubric, optional probabilities and confidence	A position on an ordered, zero-indexed rubric.
check({ instructions })	probability	The estimated probability of yes. Your code chooses the action threshold.

Every question requires non-empty instructions.

1. Choice

const department = choice({
  instructions: 'Which department should handle this?',
  options: {
    billing: 'Payments, invoices, and refunds',
    technical: 'Product bugs and technical problems',
    general: 'Other requests',
  },
})

Option keys are the labels. Option values describe each label to the model and may be any JSON value, including structured criteria or null for an undescribed label. Keys must be non-empty and at least one option is required. Because the options are captured as literals, the answer's choice is typed as the union of the keys.

2. Multiple labels

const topics = multiLabel({
  instructions: 'Select all topics present in the message.',
  options: {
    subscription: 'Subscriptions and renewals',
    duplicateCharge: 'Multiple charges for the same purchase',
    refund: 'Requests to return a payment',
  },
  threshold: 0.7,
})

Labels are judged independently, so none, one, or several can be selected. threshold is inclusive and must be between zero and one; it defaults to 0.5. Probabilities do not need to sum to one.

Providers may compose this from several native questions. With Jev, each label becomes its own question in the same request, which can increase billed usage.

3. Score

const urgency = score({
  instructions: 'How urgently does this need attention?',
  rubric: ['Low', 'Normal', 'High', 'Critical'],
})

The rubric is an ordered list of at least two JSON levels. The returned score is a position on the zero-indexed rubric and can be fractional, so 1.6 sits between Normal and High. The answer repeats the rubric you sent, preserving its literal types. To prioritize items, score them on the same rubric and sort or combine dimensions in application code.

4. Check

const cancellation = check({
  instructions: 'Is the customer asking to cancel their subscription?',
})

A check returns a probability between zero and one. Anvia does not pick a threshold for you: decide in application code what probability should trigger an action.

5. Combine questions

const questions = { department, topics, urgency, cancellation }

Question names identify the answers. A single request may mix question types when the model declares mixedQuestions. Define question objects once and reuse them across inputs; they are plain data.

6. State

state accepts any JSON-compatible value: a string, number, boolean, null, array, or object. Anything else, such as a Date, class instance, or undefined field, is rejected before the call.

Validation rules

Rule	Error
Questions must be plain data objects with non-empty instructions	TypeError
choice and multiLabel need at least one option; keys cannot be blank; values must be JSON	TypeError
score needs a rubric of at least two JSON levels	TypeError
multiLabel threshold must be between zero and one	RangeError
questions must be a non-empty plain object with non-empty names	TypeError

Next, read Answers and results.