Questions

A question is an observation. The model returns evidence. Your thresholds decide accept or review. Put the decision in review. Leave it out of instructions. “Only pick billing if you are sure” changes the answer. minSelectedProbability changes the result. Ask for something a person could see in the ticket. Ticket age, plan entitlements, counts, and date math belong in your code. Question ids are stable keys: lowercase, starting with a letter, then letters, digits, or _, at most 64 characters. Ids in one document are unique.

Choice

Mutually exclusive options. Include other when the set is not closed.

{
  "type": "choice",
  "id": "department",
  "name": "Department",
  "instructions": "Choose the team that best owns the next action for this ticket.",
  "criteria": {
    "billing": "A payment, invoice, subscription charge, or refund issue.",
    "technical": "A product defect, outage, integration, or performance issue.",
    "account": "An access, identity, security, or account-management issue.",
    "other": "The ticket does not primarily belong to billing, technical, or account."
  },
  "review": { "minSelectedProbability": 0.7 }
}

The answer is { "type": "choice", "choice": "billing", "probabilities": { ... } }. Option descriptions are what the model matches against. Option keys are what you store.

Score

Two to ten ordered levels. Each level describes a situation. “Low”, “medium”, and “high” are not situations.

{
  "type": "score",
  "id": "urgency",
  "name": "Urgency",
  "instructions": "Score the operational urgency expressed by the current situation.",
  "criteria": [
    "Routine request with no stated time pressure or active impact.",
    "Time-sensitive request or limited impact affecting one customer.",
    "Serious degradation, repeated failure, or impact affecting multiple users.",
    "Critical outage, security exposure, or immediate material business impact."
  ],
  "review": { "minInclusive": 1, "maxInclusive": 2 }
}

The answer is { "type": "score", "score": 1.4, "probabilities": { ... } }. The score is a number. The review range is inclusive on that number. Level index 0 is the first string in criteria.

Boolean

Whether a statement is true. The answer is a probability, not a hard boolean. 0.5 is the middle of that probability. It is not a third category.

{
  "type": "boolean",
  "id": "refund_requested",
  "name": "Refund requested",
  "instructions": "The customer is asking to reverse or refund a charge.",
  "criteria": {
    "true": "The customer explicitly or clearly implicitly requests money back.",
    "false": "The customer does not request money back."
  },
  "review": { "lowExclusive": 0.35, "highExclusive": 0.65 }
}

criteria is optional. When you set it, include both true and false.

Starter triage pack

A new pack is created with these five questions. Defaults: choice floor 0.7, score review range [1, 2], boolean review band (0.35, 0.65).

Id	Type	Observes
`department`	choice	`billing`, `technical`, `account`, `other`
`urgency`	score	Four situations, from routine to critical
`refund_requested`	boolean	The customer asks for money back
`policy_risk`	score	Four situations, from none to severe
`escalation`	boolean	The customer asks for a higher tier

When those ids are present, the result may include derived:

Field	Rule
`department`	The selected `department` choice
`priority`	From `urgency`: below `1` is `low`, below `2` is `medium`, otherwise `high`
`tags`	`refund` when `refund_requested` probability is at least `0.8`. `escalation` on the same rule

derived is omitted when none of those ids are answered. It is not a second policy language.

Document

A pack document is { "questions": [ ... ] } with at least one question. review is stored with the question and is not sent to the model.

Get started

Packs

Connect

Operate

Reference

Choice

Score

Boolean

Starter triage pack

Document

​Choice

​Score

​Boolean

​Starter triage pack

​Document

Choice

Score

Boolean

Starter triage pack

Document