Operations
The Operations page shows how the deployment is being used, what the AI is costing, and anything that needs attention.
| Tab | What it shows |
|---|---|
| Overview | Assessments, learners and tutor use, plus active alerts |
| Costs | Estimated OpenAI spend and token use, broken down by feature, and vector store storage |
| Acceptance | AI performance against targets |
| Deletions | Tutor conversations queued for deletion at OpenAI |
| Jobs | The background tasks and when each last ran |
Costs are estimates from token counts. Your OpenAI bill is the authoritative figure.
Budgets
Two hard limits on AI spend, both in US dollars:
| Limit | Default | Applies to |
|---|---|---|
| Per-session cap | $1 | One learner's assessment session |
| Daily global cap | $100 | All AI use across the deployment, resetting at midnight UTC |
Once a limit is reached, the tutor answers with a message that it's temporarily unavailable, and the rest of ScamSafe keeps working. Zero turns a limit off.
Alerts
Alerts appear on the Overview tab when:
| Alert | Default threshold |
|---|---|
| Error rate: the share of AI requests failing | 10% |
| p90 latency: how slowly AI requests respond | 8000 ms |
| Daily spend: today's spend against a basis you set ($100 by default) | Warns at 80%, critical at 100% |
| Provider errors: temporary OpenAI failures, such as rate limits | 3 in the window |
Zero turns an alert off. Some failures only an administrator can fix, such as an invalid API key or an exhausted OpenAI balance. These always alert on the first occurrence and can't be turned off, and the alert says what to do.
Acceptance targets
Targets for the AI's efficiency and responsiveness: how much text each tutor and assessment request sends, how often OpenAI's prompt cache is hit, how long conversations run before they are summarized, and how quickly the tutor starts answering. They are measured, not enforced: a missed target is a sign that a prompt edit or content change made requests larger or slower.
Deletions
Deleting a user also deletes their tutor conversation at OpenAI. If OpenAI can't be reached, the deletion waits here and is retried. Retry now retries everything waiting.
Jobs
Background tasks that keep the deployment healthy, such as license renewal, deleting conversations at OpenAI, and keeping corpus documents, glossary entries and personas in step with the vector stores. Run now runs one immediately.