Define what Jev should understand
Choose the criteria that Jev uses to judge the text of each request.
Product screen with sample data. Intake is the previous working name.
A Jev question has one job
A Jev question tells the model what to judge. It does not appear as an applicant question.
For support intake, Intent can distinguish Billing from Technical. In scope can test whether the request belongs to your service.
Write criteria in terms your team can test. Use examples of both matching and non-matching text.
Choose the judgment type
Pick one selects one option. Describe each option so its purpose is clear. Jev returns option probabilities and confidence.
Scale places the text on your ordered levels. Describe every level. Jev returns a weighted value, probabilities, and confidence.
Yes or no returns the chance that the answer is yes. It does not return a separate confidence value.
Do not treat a probability as a fact. Use your rules to decide when a person must review a request.
Add and test a criterion
- Add a Jev question in Understand.
- Enter its name.
- Select the type.
- Write the question and option or level descriptions.
- Enter sample text in Try this question.
- Select Try.
- Compare the judgment with what your team expected.
A live try uses the installation’s TypeSafe account.
Keep the call focused
For each accepted applicant answer, ClientStack sends all criteria in one Jev call. The input includes the applicant’s text so far.
The runtime uses the model pinned to the published version. It does not switch to another provider. The editor and replays can reuse saved judgments; a real applicant answer gets a fresh call.
Check the boundaries
Test short text, unclear text, and requests outside your service. Include text that tries to tell the model what to answer. A handoff rule is useful when confidence or scope is too low.
Read rules before you connect judgments to outcomes.