Noul
Use a yes/no question when the action depends on whether a condition holds. Read the probability of yes; set your own threshold outside the model.
JEV AI · PRODUCT GUIDE
Jev AI documentation
Design a decision, test its answers, and choose the right path from the playground to saved judges, batch work, or an API call.
START HERE
Turn one judgment into a repeatable rule
Jev reads the material you supply as a state and answers the questions you define. Start with one outcome your application needs, then test the wording against ordinary and difficult examples.
- 01
Bring the evidence
Paste a message, document excerpt, or structured record into the playground. Keep the material relevant to the decision.
- 02
Describe the boundary
Ask what must be true for each answer. For an answer supported by a source, include both the answer and the source text in the state.
- 03
Inspect uncertainty
Compare answers and probabilities across examples. Decide in your own code when to accept a result and when to send it for review.
Choose the answer your code needs
Choice
List the permitted outcomes when exactly one label should win. The response includes the selected label and a distribution across your options.
Score
Write ordered levels for a quality or risk rubric. The answer can fall between levels, so preserve the numeric score and the distribution across levels.
Need a worked example? The use case library shows decisions for retrieval, routing, evaluation, and data matching.
Move from a trial to a workflow
Save a judge
Keep a tested set of questions together, then run it on new text or call it by ID through the API.
Open My judges →Evaluate a file
Apply one judge to rows in a CSV, TXT, or JSONL file and export the results when the run finishes.
See batch processing →Check live sources
Compare a decision made with web evidence against one made without it when recency or source context matters.
Explore web context →Call from your backend
Create an API key and send a state plus typed questions to the Jev endpoint. Keep the key on your server.
Read API examples →Check the decision before you automate it
A probability is a signal to evaluate against your own examples. It is not a guarantee that any one answer is correct.
- Test ambiguous and adversarial input, not only examples that clearly fit a label.
- Keep arithmetic, exact counting, and date comparisons in code.
- Record the question version and review low-confidence results when the outcome matters.
- Check current billing and the API limits and response format before production traffic.