Dashboards

The dashboards at grafana.nufi.me show the operational side of NUFI — request rate, error rate, latency percentiles, host health.

Where the trace viewer shows what the AI said, the dashboards show how the platform held up while it was saying it.

Sign in

Use the dashboard credentials set during install. In production, front the dashboards with your reverse proxy so they share the same SSO posture as the rest of NUFI.

Pre-loaded dashboards

NUFI ships with:

• Gateway Overview — request rate, error rate, latency P50 / P95 / P99, top models by request count, top users by request count.
• Database health — connection count, slow queries, replication lag.
• Cache — hit rate, memory usage, slow operations.

All three load on first visit. Click the dashboard name in the left rail to switch.

Add a panel

Dashboards are queries against a metrics database. To add a panel:

1. Pick a dashboard → Add → Visualisation.
2. Pick the metrics datasource.
3. Write a query, e.g. sum by (model) (rate(nufi_total_requests[5m])).
4. Save.

Alerts

NUFI ships with three default alert rules:

Your operator can edit these in the alert rules file and reload without restarting the metrics service.

Routing alerts to your incident channel

By default, alerts route to a no-op receiver — they are recorded but no notification is sent. To wire Slack, Teams, or PagerDuty, ask your operator to configure the alert routing.

Retention

Metrics retention is 15 days by default. Past that, you have raw counts in the metrics database but not per-second resolution. If you need longer history, ask your operator to extend retention.

When to look here vs the trace viewer

• Dashboards — is the platform up? Are we slow? Is there an error storm right now?
• Trace viewer — what exactly did the AI see and produce for user X at time T?

You usually start in the dashboards (you noticed something), then jump to the trace viewer (to inspect a representative request).