Our summary
A support message can raise several questions: which team should receive it, does the customer want a refund and how serious is the problem? Piotr Płoński shows how Python software can ask Jev for these judgments without asking it to write a reply.
One question type selects a named option, another estimates the chance of yes, and a third places the message on a described scale. Several questions can share the same message in one request. The program still owns the next step; the example action functions stand in for work you must implement.
Answers that fit your allowed options can still be wrong. Płoński recommends testing difficult messages before choosing when to trust them. His comparison with an OpenAI model records only one call with different prompts and unequal timing setup, so it cannot establish a general speed or accuracy ranking.
Key takeaways
- Use named options for categories, a yes probability for yes-or-no questions and described levels for ratings.
- Ask independent questions about the same information together, then let your own program decide what happens next.
- Test ambiguous examples and failed requests. One successful comparison is not a repeatable performance benchmark.