The essentialsSources checked October 14, 2026
What Is Jev AI?
Jev AI is TypeSafe AI’s model for making structured decisions from text. Give it information and a question, and it returns a category, a score, or a yes-or-no probability. It can help sort messages and assess inputs, but it does not write replies. The examples below show what to ask—and how to tell whether the answers are useful.
What does Jev AI return?
Jev supports three question types:
| Question type | What you ask | What the result represents |
|---|---|---|
| Choice | Which category fits this input? | A selected option and probabilities across the options |
| Score | Where does this input fall on a defined scale? | A score and a distribution across the scale |
| Noul | Is this statement true? | A probability between 0 and 1 |
Choice and Score also include a separate confidence value. Noul expresses uncertainty through its probability rather than a separate confidence field. TypeSafe’s question types · Confidence documentation
Imagine a customer writes:
“The export button stopped working after yesterday’s update. I need the report for a meeting this afternoon.”
You could ask Jev to choose a department, rate the impact, and assess whether the customer mentions a deadline. These should remain separate questions: an approaching deadline and a broken feature tell you different things.
This example illustrates how to frame a request; it is not a recorded model result.
How is Jev AI different from a chatbot?
A chatbot can write a response, explain a concept, or draft a report. Jev AI returns answers within the formats you define. It does not write the customer’s reply or explain its reasoning in a paragraph.
The two can work together. Jev could help select the appropriate workflow, while a text-generation model drafts a response using the relevant information. Your application still controls which actions are allowed. TypeSafe’s System One guide
Where does Jev AI fit?
A useful starting point is a repeated judgment with a clear destination.
For example, suppose your team manually labels incoming feedback. Before involving a model, write down what counts as a bug report, a feature request, or a billing question. Then collect messages that are easy to classify and messages that sit between categories.
This gives you something concrete to test. “Does Jev agree with our labels?” is a more useful question than “Is Jev intelligent?”
Other possible experiments include checking whether a retrieved passage answers a question or assessing a message against a clearly defined review policy. Begin with one decision that you can inspect.
Does a valid answer mean a correct answer?
No. An answer can use the correct format and still select the wrong category.
Jev’s probabilities help express uncertainty, but they do not guarantee that an individual judgment is correct. Calibration concerns how predictions behave across a set of examples. TypeSafe’s explanation of calibration
When reviewing results, look at the mistakes as well as the successes. Did the input omit a necessary fact? Were two categories nearly identical? Did your question ask the model to infer something the text never stated?
How should you try Jev AI?
Start with a small collection of examples you understand. Define the answers before testing, keep a few examples aside, and check those after adjusting your questions.
If you want to explore the request format in a browser, try Jev on UseJevAI. UseJevAI is an independent website, separate from TypeSafe AI.