Independent guideAn independent guide to Jev by TypeSafe AI
Use cases

Ask only the question that matters next

Build dynamic forms and workflows with Jev

Let Jev pick the next question or branch from the answers so far, while your form keeps the fields, validation and final decision.

Load this template

Sources reviewed Updated Content revision 1

Evidence Original fictional teaching example. No model result is prefilled.

Original fictional teaching example. No model result is prefilled.

An editable starting point

Load this template
1 · Text to check
Form: equipment support request Answers so far: - Device: laptop - Problem: “The screen flickers when I move the lid. It started after I dropped it.” - Purchased: 14 months ago
2 · Your question
Which follow-up should the form ask next?
3 · Define 2–6 options
  1. Ask for photos of the damage

    The answers describe physical damage that a picture would document.

  2. Ask which troubleshooting steps were tried

    The problem sounds like software or settings.

  3. Ask about warranty or a protection plan

    Coverage decides the path and is not yet known.

  4. No more questions; send for review

    The answers already determine the next step.

These are model probabilities, not measured accuracy or a guarantee. Yes and No describe your rules; choices are relative to the options you supplied.

Can Jev decide the next step of a form?

Yes, when the possible next steps are already written down. A static form branches on exact values: if the device is a laptop, show the laptop section. Free-text answers break that logic, because “it started after I dropped it” matches no dropdown value. A Choice question can read the answers so far and select one of the follow-ups you prepared. Your form still owns the fields, the validation and whatever happens after submission.

Open the editable next-question template. It contains a fictional equipment support request with three answers and four possible follow-ups. Loading it creates a draft; a check starts only when you request it.

Write follow-ups as options with reasons

Follow-up Choose it when
Ask for photos of the damage The answers describe physical damage that a picture would document.
Ask which troubleshooting steps were tried The problem sounds like software or settings.
Ask about warranty or a protection plan Coverage decides the path and is not yet known.
No more questions; send for review The answers already determine the next step.

Always include a way to stop. Without it, the form keeps asking because some option must win. Keep required legal or safety questions out of the model’s hands: if a field is mandatory, ask it in code every time.

Change one thing and compare

  1. Run the template and note the selected follow-up and the other probabilities.
  2. Replace the problem with “The screen flickers after the latest system update” and run it again.
  3. Add a line “Protection plan: active until 2027” and see whether the warranty question loses its purpose.

No expected answer is prefilled, and a few runs are not a test of your form. Before launch, collect real or realistic partial submissions, write down the follow-up a trained colleague would ask, and compare.

Ask several things in one request

In your own integration, one request can carry many questions. TypeSafe’s speculative fan-out pattern sends the likely questions together, including ones that may turn out to be irrelevant, and lets code read only the answers that apply to the chosen branch. Its documentation says the questions are evaluated in parallel, so adding more usually has little effect on response time. For a form, that means one call can return the category, whether an order number was given and whether the tone calls for a callback, and the form decides what to show. The Playground on this site asks one question per check, so it demonstrates a single step of that design.

What the community has tried

A password-manager team shared on 17 September 2026 an experiment that uses Jev to recognise fields in complex web forms, replacing hand-written rules. The improvement and latency figures in that post are the authors’ own, from their own test pages; they have not been independently reproduced and were described as an experiment, not a full deployment. More examples are in the Projects gallery.

A model can misread sarcasm, mixed languages or an answer typed into the wrong field. Give people a visible way to reach every branch by hand, and read where Jev’s limits are before shipping.

Sources & further reading

Official documentation and linked community reports support this guide. Community observations are attributed to their authors.

  1. TypeSafe: Speculative fan-out
  2. TypeSafe: How to build with TypeSafe
  3. TypeSafe: Choice
How we check our sources