| Mise à jour | 29 sep. 2026 04:40:11 | 29 sep. 2026 04:40:13 | 29 sep. 2026 04:40:15 |
|---|
| Site web | jev-ai.pro [Jev API] | jev-ai.pro/... [Jev API documentation] | jev-ai.pro/... [Jev API documentation] |
|---|
| When to use | Select one label from a fixed, unordered set. Useful for routing
tickets to billing, technical, or account queues. | Evaluate an input against an ordered rubric. Useful for urgency
levels with explicit definitions for each score. | Estimate whether one clearly stated proposition is true. Useful for
a separate escalation flag, independent of the routing label. |
|---|
| Example question | Which team should handle this ticket? Choices: billing, technical,
account. Include a fallback policy for unclear inputs. | How urgent is this ticket? Define a rubric such as 1 = routine
question, 2 = degraded function, 3 = service blocked. These are
illustrative business definitions. | Does this ticket describe an active account-security incident?
Phrase one testable proposition rather than combining unrelated
conditions. |
|---|
| Result to inspect | choice plus confidence and probabilities. Validate the label
against the allowed choices; do not assume confidence equals the
winning label probability. | score plus confidence, legend, and probabilities. Interpret the
score using the declared rubric, not as a universal measure of
severity. | noul: a value from 0 to 1 representing the model's estimated truth
probability. It is an estimate, not verified evidence that an
incident occurred. |
|---|
| Human review boundary | Route unknown labels, malformed responses, and ambiguous cases to
review. Keep refunds, deletions, and access changes behind separate
authorization. | Choose action thresholds using a labeled evaluation set and the
cost of mistakes. A high urgency score can trigger review without
directly executing an action. | Use a review band between positive and negative decisions. Set
thresholds with representative cases; do not copy a numeric
threshold as a default safety guarantee. |
|---|
| Evaluation approach | Measure per-label precision and recall, plus confusion between
queues. Include empty text, mixed intents, unseen topics, and
instructions embedded in customer text. | Measure disagreements against human rubric labels, especially large
errors. Review whether two annotators interpret each level
consistently. | Measure false positives, false negatives, and calibration on
labeled statements. Recheck results when input language or the
incident definition changes. |
|---|
Avis utilisateurs et commentaires