Alerts
Alert metrics
The nine metrics an alert can watch: what each measures, its unit and default threshold, which use minimum events, the windows and the channels.
Reference2 min readUpdated 31 October 2026
Each alert watches one of these metrics. The default threshold is set when you pick the metric; change it to suit your traffic.
Metrics
| Metric | What it measures | Unit | Default | Minimum events |
|---|---|---|---|---|
| Error rate | Share of replies that failed | % | Above 5 | Yes |
| Errors | Replies that failed | Count | Above 10 | No |
| p95 latency | 95% of replies finished faster than this | ms | Above 15,000 | Yes |
| Requests | Replies your agents gave, successful or not. Use Below to catch traffic stopping | Count | Below 1 | No |
| Thumbs-down rate | Share of rated replies that got a thumbs down | % | Above 20 | Yes (ratings) |
| Average score | The average of one score key, shown 0 to 100 | Of 100 | Below 70 | Yes (scores) |
| Credits spent | Credits used in the window | Credits | Above 500 | No |
| Credits left | Your balance right now. No window and no channels | Credits | Below 100 | No |
| Workflow failures | Workflow runs that failed | Count | Above 0 | No |
How they're measured
- A reply is one answered request. Classify, Extract and Guardrail check calls inside workflows count on the API channel.
- p95 latency, thumbs-down rate and average score use the newest 5,000 replies, ratings or scores in the window.
- Average score is a score key's average times 100, so scores of 0.7 read as 70. Experiment results are left out. The key is an evaluator's, a review queue rubric's, or one you send with
POST /v1/feedback. See Feedback and scores. - Credits spent for one agent counts its replies (1 credit each). For All agents it also counts your account's other spending: knowledge, the workflow assistant, Insights reports and evaluations.
- Workflow failures count failed runs started in the window. The designer's test runs count only when Playground is ticked.
Windows and cooldowns
| Setting | Choices | Default |
|---|---|---|
| Over the last | 5, 15 or 30 minutes; 1, 6, 12 or 24 hours; 3 or 7 days | 1 hour |
| Don't repeat within | 5, 15 or 30 minutes; 1, 3, 6 or 24 hours | 1 hour |
| Minimum events | 0 to 100,000 | 5 |
Channels: Live and Test
| Channel | Environment | Traffic |
|---|---|---|
| Widget | Live | Your website's chat widget |
| Chat link | Live | Hosted chat links |
| API | Live | Your own code, and workflow steps that call agents |
| Playground | Test | The Playground and an agent's test chat; for workflows, the designer's test runs |
New alerts watch the Live channels, so testing in the Playground never pages anyone. Tick Playground to include test traffic.
Still stuck?
Email us with what you tried and any request id. A person replies.