Submit a projectJev evaluates a state against small, defined questions and returns typed judgments. It fits inside a larger workflow: choosing among candidates, applying a rubric, or evaluating a statement.
Three useful primitives
- Choice: select an option and receive a probability for each candidate.
- Score: evaluate content against an ordered rubric.
- Noul: estimate the probability that a statement is true.
You can ask several independent questions in one request. One answer does not become the context for another; code must handle dependencies between requests.
Where to try it first
Shortlisting search results, routing support requests, and selecting from observed browser controls are practical starting points. Keep exact rules, arithmetic, date comparisons and authorization in code.
Know the boundaries
Jev currently accepts text and JSON, rather than images, audio or video. It does not generate explanations or code. A schema-valid answer can still be wrong, and confidence describes the distribution rather than guaranteeing an individual decision.
The stable model checked for this guide is jev-1.13.0. Versions, pricing and limits can change; check the provider before integrating.
Make a small experiment
Choose one narrow decision. Prepare labeled examples, including missing targets and insufficient evidence. Then test whether the answers actually help your workflow.