jevA support message is a useful example
A customer writes, “My delivery is three days late.” Your application needs a team name: shipping, billing or returns. That is the kind of decision Jev is designed to make. You supply the message and describe the allowed categories; Jev returns a structured answer.
The three parts of a request
The state is the information to inspect. The questions describe what you want to know. The answers use the same question identifiers, so your code can find the result. In the example below, department is an identifier; the actual instruction is written inside that question.
Use it for decisions that have defined outputs
Other examples include scoring a retrieved paragraph or checking whether a message asks for urgent help. If the result must be a written reply, you still need a text-generating model. Jev can help choose the handler before that reply is written.
Try one task before changing your application
Start with messages you can label yourself. Include an unclear message and one that mentions two problems. Compare the resulting decisions with your existing rules. The interesting question is whether the extra model call fixes mistakes that matter to your users.
Request example
{
"model": "jev-latest",
"state": "My delivery is three days late. Can you help?",
"questions": {
"department": {
"type": "choice",
"instructions": "Which support team should handle this message?",
"criteria": {
"shipping": "Delivery status and tracking",
"billing": "Invoices and payment problems",
"returns": "Refunds and exchanges"
}
}
}
}