Use Choice to route an input to one of 1–26 options. Add an other or unclear option when your list may not cover the input.
json
{
"team": {
"type": "choice",
"instructions": "Which team owns the ticket?",
"criteria": {
"billing": "Payment issues",
"shipping": "Delivery issues",
"other": "No matching team"
}
}
}An illustrative answer:
json
{
"type": "choice",
"choice": "billing",
"confidence": 0.7,
"probabilities": {"billing": 0.8, "shipping": 0.1, "other": 0.1}
}Confidence is (top_probability - 1/n) / (1 - 1/n). It compares the winner with a uniform distribution — it is not the measured probability that the answer is correct.
Criteria values can be descriptions, structured JSON, or null when a key is clear on its own. Test descriptions on your own cases. A single option returns probability 1 without inference, so it cannot test model quality.