Jev is for decisions, not conversations
A general language model is optimized to continue text. Jev is useful when an application needs a bounded answer such as a queue, risk level, approval signal, or next action. The output shape is part of the request, so downstream code has less interpretation work.
State is the evidence available now
State is the compact set of facts Jev should consider: a support ticket, a lead record, a moderation event, or the current step in an agent workflow. Good state is relevant, inspectable, and free of fields that cannot change the decision.
Questions define the contract
Use Choice for one option from a named set, Score for an ordered level, and Noul for a bounded yes-or-no style decision. Start with one question so you can see whether the criteria are mutually exclusive and useful.
Confidence only matters with a fallback
A confidence value becomes operational when your software has a threshold and a documented action below it. Low-confidence decisions can enter human review, ask for more state, or use a safer default instead of silently continuing.
Jev questions beginners ask
Who created Jev?
Jev was introduced by TypeSafe AI as a model for structured software decisions.
Is Jev a chatbot model?
It can interpret natural-language state, but its main use is returning bounded, typed decisions rather than holding an open-ended conversation.
What can Jev return?
The core patterns are Choice for a category, Score for an ordered value, and Noul for a bounded binary-style decision, typically with confidence information.
When should I use Jev instead of an LLM prompt?
Use Jev when the next software action depends on a small, explicit decision and you want the expected output shape visible in the request.
Does confidence guarantee correctness?
No. Confidence is a control signal, not proof. Validate it on your own examples and define a fallback for uncertain or provider-error cases.
What is the easiest first Jev project?
Choose one observable classification or routing task, keep the criteria small, and review every result until the threshold and fallback are defensible.