Independent guideAn independent guide to Jev by TypeSafe AI
Get started

Make category boundaries explicit

Handle overlapping labels in Jev classification

Separate competing routes from coexisting properties, define category boundaries, and keep a review path for mixed messages.

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
I was charged twice for my monthly subscription. Can someone help me check the two payments?
2 · Your question
What kind of feedback is this?
3 · Define 2–6 options
  1. Bug or fault
  2. Feature request
  3. Billing
  4. Help with usage
  5. Positive feedback
  6. Other

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

Name the conflict

Overlapping labels may reveal two different problems. Your workflow might need one route but have vague definitions, or the input may genuinely have several properties. Rewording a category cannot make a two-issue message contain only one issue.

This tutorial uses an original feedback template. It does not reproduce a community test or establish a best-performing set of labels.

Edit the rules, not just the names

For one-owner routing, specify what each label covers and excludes, and how to resolve a message that fits two labels. Choose a policy that your support team can explain. Add an out-of-scope or review option where needed; such an option gives the model a possible answer, not a guarantee that uncertainty will be detected.

If the goal is to record all applicable properties, use separate yes/no checks in the wider workflow. The custom editor lets you explore one such check at a time. It does not turn a Choice question into multi-label output.

Inspect mixed and missing information

Use a fictional message with a payment problem and a login failure. Compare it with two messages containing only one issue each. Then try a vague complaint with no actionable detail. Decide the intended handling before running, and retain mistakes as examples to revisit after a wording change.

A leading label can still be wrong. Close probabilities may prompt review, but a confident result is not evidence that an omitted category was unnecessary.

Choose the next step deliberately

For low-information messages, your application may ask for clarification or send them to a person. It should not invent account facts or treat the proposed label as permission to refund. No universal probability cutoff is established by this tutorial.

Return to the feedback recipe, or compare Choice and Noul when deciding whether you need one route or several independent properties.

Sources & further reading

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

  1. TypeSafe: Choice
  2. TypeSafe: Noul
How we check our sources