Metrics Enterprise¶
Your operations team already watches everything else in one monitoring stack; the AI gateway shouldn't be the one thing they can't see. The gateway exports its own metrics in the Prometheus text format, so request rates, latency, tokens, errors and provider health go into the dashboards and alerts you already run.
What Enterprise includes depends on the features your licence includes. On Community, the Control room's dashboards, Request logs and Status (Check that Vidai is working) show traffic and health inside Vidai.
Where to read them¶
- The gateway serves
GET /metricson a port of its own (9091 unless your installation sets another), apart from the address your apps call. Metrics are off until the gateway's configuration switches them on. - The same port answers
GET /healthwith{"status":"healthy"}, for a liveness probe. - The metrics port asks for no credentials. Keep it on your own network, or put it behind your own access control.
- Each gateway node serves its own metrics; your monitoring adds them up across nodes.
- The metrics are registered when the gateway starts, so the first scrape already finds them.
The metrics¶
| Metric | Type | Labels | What it counts |
|---|---|---|---|
vidai_requests_total |
counter | provider, model, status_code |
Calls, by the status the caller got |
vidai_request_duration_seconds |
histogram | provider, model |
Time from call to answer (buckets from 0.1 s to 60 s) |
vidai_queue_duration_seconds |
histogram | provider |
Time a call waited before it was sent (buckets from 1 ms to 5 s) |
vidai_active_requests |
gauge | — | Calls in progress now |
vidai_active_requests_by_provider |
gauge | provider |
Calls in progress now, per provider |
vidai_prompt_tokens_total |
counter | provider, model |
Input tokens, as the gateway counts them |
vidai_completion_tokens_total |
counter | provider, model |
Output tokens, as the gateway counts them |
vidai_vendor_input_tokens_total |
counter | provider, model |
Input tokens, as the provider reports them |
vidai_vendor_output_tokens_total |
counter | provider, model |
Output tokens, as the provider reports them |
vidai_cached_tokens_total |
counter | provider, model |
Input tokens the provider served from its prompt cache |
vidai_reasoning_tokens_total |
counter | provider, model |
Reasoning tokens, where the provider reports them separately (OpenAI's reasoning models) |
vidai_cost_dollars_total |
counter | provider, model, currency |
Cost as reported by a provider that reports cost per call (OpenRouter) |
vidai_errors_total |
counter | error_type, provider |
Failed calls: unauthorized (401), forbidden (403), rate_limited (429), server_error (5xx), client_error (other 4xx) |
vidai_rate_limit_hits_total |
counter | provider |
Calls answered 429 |
vidai_provider_up |
gauge | provider |
1 while the provider is taking calls, 0 when it isn't |
Good to know¶
- Cost in Vidai's own currency, priced from the rate card, isn't a metric: it's in Cost, and you can send it to your own storage with Send data to your BI tools.
- The gateway records metrics as it writes each batch of calls, never on a call's way through, so a slow monitoring system never slows a call.
Related¶
- The parts: what runs where
- Check that Vidai is working
- Read provider health and circuits