Jev Choice vs Score vs Noul: which question type should you use?
A practical guide to TypeSafe AI's three Jev question types, with examples of categorical choices, ordered scores, and yes/no probabilities.
Start with the decision your code needs
Jev's questions are typed: you define the answer shape before sending the input. The three types are Choice, Score, and Noul. Pick the one whose answer your application can act on without trying to parse prose.
TypeSafe AI recommends one focused judgment per question. If a decision depends on several factors, ask separate questions and combine their answers in your own code. All questions in one request see the same state.
Choice: select from named alternatives
Use Choice when the result should be one of a known set of labels: accounts, billing, or other; bug, feature request, or question. Define the options and what each means. The returned choice is a label your routing code can recognize. The answer also contains option probabilities and a confidence value.
Make the alternatives distinct and include a sensible fallback where needed. A vague label set can produce an apparently valid answer that is still hard to use.
Score: pick a level on an ordered scale
Use Score when the levels have an order, such as calm, frustrated, and very angry. Its criteria are an ordered list. The answer includes a score, a legend mapping levels to their descriptions, probabilities, and confidence.
Do not use Score just because you want a number. First define what each level means and how a human would distinguish adjacent levels. Otherwise the number will be difficult to interpret or evaluate.
Noul: judge whether one statement is true
Use Noul for a focused yes/no proposition: Does this message ask for a refund? Does it express urgency? The returned noul value lies between 0 and 1. It represents the model's judgment of the proposition, so an answer near the middle signals ambiguity for that question.
Noul does not return a separate confidence field in the documented response. Avoid treating its value as an extra confidence score attached to a binary label. If your app needs a yes/no action, choose a threshold based on mistakes you can tolerate and test it on your own data.
A support-routing example
For one customer message, ask Choice for the responsible queue, Noul for an explicit refund request, and Score for frustration level. Then write ordinary code that decides which queue receives the message and when a person should review it. This keeps the business rule visible and lets you adjust it without rewriting a broad prompt.
You can see a working Python request in our Jev API quickstart. Run it on messages from your own workflow before trusting any automatic action.
Source: TypeSafe AI question primitives ↗. This directory is independent and its explanations are editorial summaries, not TypeSafe AI documentation.
Related: Try the Python quickstart.