Jev: a decision model, not a chat model
Jev isn’t a chat model — it’s a decision model for code.
Instead of a free-form conversation, you ask a question, it answers with probabilities. You give it a list of choices, it picks one.
For example, a customer writes:
"I was charged twice. Please fix this ASAP."
You want to know if it’s a dispute:
noul: "Is this a billing dispute?"; {true: "The customer disputes a charge."; false: "Something else."}
Then categorize it:
choice: "What is this ticket about?"; {billing: Null; technical: Null; other: Null}
And how urgent it is:
score: "How urgent is this ticket?"; ["Not urgent."; "Somewhat urgent."; "Needs attention today."]
The response looks like this:
"billing": { "type": "noul", "noul": 0.98 },
"category": {
"type": "choice",
"choice": "billing",
"confidence": 1,
"probabilities": { "other": 0, "technical": 0, "billing": 1 }
},
"urgency": {
"type": "score",
"score": 1.96,
"confidence": 0.93,
"legend": {
"0": "Not urgent at all.",
"1": "Somewhat urgent.",
"2": "Needs attention today."
},
"probabilities": { "0": 0, "1": 0.04, "2": 0.96 }
}
From there, your code can escalate the ticket, create a task, tag it with a category, and so on.
It’s fast and cheaper than a chat model — and with the structured, typed response, you can plug the result straight into your own logic instead of parsing free text.
Jev with 4D
typesafe-sdk-4d brings Jev to 4D.
A lot of things could be built on top of this (some of what’s floating around online is “fake” — it’s a new buzzword), but paired with a standard LLM, it can do some genuinely useful work.
Two curated lists of resources: awesome-jev and awesome-jev — use cases, projects, SDKs, and more.
Find the full 4D component catalog on the home page.