Choice: one option
Choice returns the selected category, a probability distribution and confidence. The documentation sets a maximum of 255 options. The meaning of the task belongs in the instructions and the option descriptions; the question’s technical identifier is not sent to the model. S03
Our proposed categories for customer support: complaint, shipping, invoicing and other. Add edge cases to the descriptions. “Other” should be a genuine option, not a dumping ground for vaguely written rules.
Score: a position on a scale
Score uses ordered levels described in words, at most ten. The result is the probability-weighted expected value of the level index; it can be a decimal. The probabilities and confidence are included. S04
An illustrative relevance scale: 0 = unrelated, 1 = marginally related, 2 = strongly related, 3 = directly answers the question. A value of 2.4 is a rating on this scale, not a physical measurement.
Noul: the probability of yes
Noul answers a single statement with a number from 0 to 1. It has no separate confidence field. A value of 0.8 is an estimate of the probability of a positive answer, not an “intensity of eight points.” S05
How to design your own questions
Start with one specific thing. “Is the customer asking for a replacement?” is easier to check than “What all should we do about this?” Walk through the positive and negative branches on examples in advance. Handle dependencies between questions explicitly in the application.
The question builder on JEV.sk creates JSON for the selected type. It does not call the model and does not guarantee that your business meaning has been captured correctly.