Jev Choice

Choice is the primitive you reach for when the product already has a closed set of destinations: teams, intents, document types, model names.

Request

{
  "department": {
    "type": "choice",
    "instructions": "Which team should handle this?",
    "criteria": {
      "billing": "Payments, invoicing, refunds",
      "technical": "Bugs, outages, integrations",
      "sales": "Pricing, upgrades, new accounts",
      "other": "None of the above"
    }
  }
}

criteria is a map. Values may be null when the key is self-explanatory. Official cap: 255 options.

Response

{
  "type": "choice",
  "choice": "billing",
  "probabilities": { "billing": 0.88, "technical": 0.12, "sales": 0.0, "other": 0.0 },
  "confidence": 0.81
}

Official confidence docs approximate three-option confidence as (3 × largest probability − 1) / 2. You can compute your own statistic from probabilities.

How to write criteria

Describe membership, not vibes.

• Good: “Charges, invoices, refunds, subscriptions”
• Weak: “Billing-related stuff”

When two options neighbor each other, write the boundary. Test those sentences against labeled examples.

Question IDs are not sent to the model. Put the full question in instructions.

Workflow

ticket text
   → Choice(department)
   → if confidence < T: human
   → else: enqueue(department.choice)

See customer support routing.

When to use Choice

Routing, intent, language-of-this-file, which tool to call, which skill to suggest.

When not to use Choice

• Measuring intensity → Score
• A single yes/no probability → Noul
• Open-ended extraction → regex or an LLM, then Choice over candidates

Common mistakes

• No other bucket, so every weird ticket becomes “technical”.
• Options that overlap so badly confidence collapses.
• Treating choice === "billing" as safe while ignoring confidence.
• Expecting a Choice yes/no to match a Noul on the same sentence. Official jaggedness notes say they are not interchangeable.

Python and TypeScript helpers: Python tutorial, TypeScript tutorial.