When should you use Jev instead of a text-generating model?

A useful starting point is the shape of the answer your application needs. If the next step is choosing a handler, evaluating a condition, or scoring a defined rubric, Jev is worth evaluating. If the next step is writing a reply, producing an explanation, or solving a problem through several reasoning steps, a generative model may fit better.

TypeSafe describes Jev as a model for structured judgments. You supply context and typed questions; your code receives bounded answers rather than a paragraph to interpret. That is a description of its interface, not evidence that it will outperform another model on your task. See the official introduction.

Start with the action after the answer

Imagine a support inbox. There are at least three different jobs:

These jobs can share a workflow. The routing decision does not need to compose the reply, and neither model should calculate a refund that your billing code can determine exactly.

Before choosing a model, write down the action that consumes its output. If you cannot name that action, your question may still be too broad.

A bounded answer needs a bounded question

“Handle this customer” is not a useful decision contract. It mixes intent, urgency, policy, permissions, and communication.

Instead, ask separately which department fits and whether the customer describes an urgent deadline. Your application can then decide whether a specialist or a person should handle the message. Our support-routing example includes the question, possible teams, and a fallback.

Bounded outputs make integration easier, but a valid team name can still be the wrong team. Retain a review path and evaluate examples outside the expected categories.

Keep exact operations in code

Use code for arithmetic, date comparisons, authorization, and database lookups whenever the operation is precisely specified. A semantic judgment can inform an action; it does not grant permission to perform it.

For example, detecting that a customer wants an account closed is different from confirming their identity and executing the closure. Those latter steps belong to the surrounding application.

TypeSafe’s version-specific notes for Jev 1.13 document weaknesses including numerical work and multi-step indirection. Treat those notes as constraints to test, not a permanent description of every future release.

Run a small comparison before committing

Take representative inputs from the job you actually need done. Label the desired outcomes. Compare a simple code baseline, your current approach, and Jev on the same cases.

Record incorrect actions, fallback frequency, latency, and total cost per completed workflow. A fast first call is not enough if its mistakes or fallbacks make the full workflow worse.

For an initial experiment, choose one narrow decision whose result a developer can inspect. Screening a diff for a possible runtime bug is one example: the output guides a review rather than claiming to prove a defect.