Learn how to use the TypeSafe AI JEV model & Question types (noul, choice, score)
Farzam Arefi demonstrates using the TypeSafe AI playground to configure Jev primitives. The video breaks down state evaluation and details how to format Noul, choice, and score question types, interpret probability distributions and confidence scores, and structure multi-primitive parallel requests.
Original by Farzam ArefiGetting startedBeginner15 min 9 secPublished
The playground separates the state being evaluated from typed questions, instructions, and criteria.
Choice returns a selection and probability distribution over the supplied criteria; a valid response can still be a misclassification.
The Score example uses an ordered criteria list and returns a fractional position across that rubric.
Worth knowing
Constraining outputs to predefined criteria eliminates text hallucination but does not guarantee the model will select the correct decision or classification.