OpenNVR surfaces its live health in a few places rather than one giant dashboard: a System & Network Monitoring card on the Dashboard (network-security alerts), and the AI Adapters page, which doubles as the fleet monitor. This page explains what each shows and where to go for deeper telemetry.

[!NOTE] The Dashboard is always visible. The AI Adapters fleet view (AI & Automation → AI Adapters) requires the ai.view permission.

What you can do here

• See at a glance how many network-security alerts fired and how severe they are.
• Jump straight from an alert tile into Alerts & Incidents filtered to that slice.
• Read AI adapter fleet health — p95 latency, request rate, saturation, and model drift — in one strip.
• Drill into any adapter for latency percentiles, outcome splits, queue depth, hardware (CPU/GPU/VRAM), and one-hour trend sparklines.

Network-security monitoring (Dashboard)

The System & Network Monitoring card on the Dashboard summarizes Suricata IDS activity. Three tiles across the top are clickable and take you to the matching view in Alerts & Incidents:

Tile	Shows	Opens
Total Alerts	Count of all alerts	All alerts
High Severity	Count of severity-High alerts	High-severity alerts
Top Category	The most frequent alert category	That category’s alerts

Below the tiles, two charts give the trend and the mix:

• Alerts over time — an area chart of alert counts.
• Severity distribution — a bar chart split into High, Medium, Low, and Unknown.

[!TIP] For raw system and service logs, see the Logs & Forensics page (Security & Network → Logs & Forensics).

AI Adapter Fleet Monitoring

The AI Adapters page doubles as the fleet monitor. KAI-C — the sovereignty and audit gateway — scrapes every adapter’s /metrics on its 60-second registry poll into a one-hour rollup, and the page renders it decision-first.

• Fleet strip — the five-second answer before any scrolling: N/M adapters ok, worst p95 (named), fleet-wide req/min, adapters near capacity, and adapters with model drift (1h).
• Per-adapter metrics (expand Metrics on a card) — each panel is captioned with the decision it drives:
  • Inference latency — p50 / p95 / p99 bars.
  • Outcomes — split across ok, model_error, and transport/provider errors.
  • Saturation — inflight vs. the declared ceiling (max_inflight); flags at ≥ 80%.
  • Queue depth — frames waiting on the model.
  • Fingerprint / drift — count and timestamp of model weight changes.
  • Trends — last hour — sparklines for p95, req/min, and inflight.

GPU & VRAM saturation

GPU and memory pressure shows up in the adapter’s Hardware panel (inside Metrics). Adapters may export optional gauges — adapter_process_cpu_percent, adapter_process_memory_bytes, adapter_gpu_utilization, and adapter_gpu_memory_bytes — and when present the panel shows CPU, Memory, GPU util, and GPU mem plus a CPU trend sparkline. If the adapter doesn’t export them, the panel says so rather than guessing.

[!TIP] Use the Hardware panel to catch VRAM pressure before you load a model that would exceed the card. The decision the panel drives: smaller model / bigger card / move the adapter.

Long-term history

Retention is deliberately delegated. Adapters speak native Prometheus exposition on their /metrics endpoints, so point Prometheus + Grafana at them for weeks-scale history. The in-app fleet view is intentionally a rolling one-hour window.

Network interfaces

The Dashboard card covers alerts, not interface throughput. To configure the Camera LAN and Uplink NICs — interface name, DHCP, MTU, IP addressing, scan subnets — use the Network page. See Network.

📸 Screenshots to capture

Save these to opennvr-site/public/screenshots/ with the exact filenames below:

• system-monitoring-fleet-strip.png — the AI Adapters Fleet strip showing adapters-ok, worst p95, fleet req/min, near-capacity, and drift counts.
• system-monitoring-adapter-metrics.png — one adapter card with Metrics expanded, showing the latency, outcomes, saturation, queue-depth, and trend panels.