High confidence
Auto-suggest the Choice (tag, route, or verdict) in your UI.
Prompt craft
How to write TypeSafe Jev questions
Jev answers the questions you declare — it does not invent your taxonomy. Good criteria beat longer prose. Keep option names stable; put nuance in descriptions.
Choice vs Score vs Boolean
| Type | You supply | You receive | Use when |
|---|---|---|---|
choice | Named options → descriptions | Selected key + per-option probabilities (+ confidence) | Mutually exclusive labels / routes |
score | Ordered levels (2–10), low → high | Probability-weighted level score + rung probs | Severity, quality, risk rubrics |
boolean | Yes/no proposition (+ optional criteria) | Probability of true | Single fact checks |
Score is not accuracy. A score of 1.5 on a 0–2 rubric is an expected level, not “75% correct.” Do not treat Score or Choice confidence as an accuracy promise.
Writing strong criteria
- Disjoint options. If two Choice keys can both be true, operators will fight the model.
- Describe edges. Say what buy_intent includes (stock, ship-to) vs price_question.
- Keep residual last. Use other / human as escape hatches, not lazy catch-alls.
- Order Score ascending. Criteria arrays are lowest → highest.
- One decision per question. Split “route” from “urgency” into separate keys.
Shared confidence policy
TypeSafe exposes confidence from the option distribution. Pack convention:
Low confidence
Send to a human queue with no default action.
Calibrate cutoffs on labeled examples from your stream or inbox. Thresholds are use-case-specific. Never wire suggestions into money or refund side effects.
State tips
Pass a string, object, or array as state. Prefer structured records (comment text + metadata) over dumping a whole chat log when only one message matters.