Our summary
Perplexity’s Decisions service answers set questions rather than writing replies. Its quickstart uses a review about broken headphones: does it report a defect, is the tone positive, mixed or negative, and how serious is the problem? This uses Perplexity’s own decision model, not TypeSafe’s Jev.
The program sends the review and three questions together. It receives probabilities for a defect, possible tones and ordered severity levels. The severity score averages the numbered levels, giving more weight to those the model thinks are more likely. Your code decides what action, if any, follows.
The service needs a Perplexity access key and receives the supplied text or images. Confidence means the model’s own estimate of how sure it is, not proof of correctness or the winning answer’s probability. Repeated answers can vary slightly, so test the numbers your code uses to decide when to act.
Key takeaways
- Ask separate questions about one piece of content together; the response keeps the names you gave each question.
- Read the full spread of probabilities and the severity scale; an average and a confidence estimate are different numbers.
- Keep secrets out of submitted content, check failures before reading answers, and test the numbers that trigger an action with real examples.