Alerts & Notifications
Get notified about errors, slow responses, outages, budget thresholds, and daily usage digests
Last updated
Alerts watch your traffic and spend, and notify you through the channels you choose. Configure them on the Alerts page (/alerts).
Alert types
| Alert | Fires when | How often it's checked | Status |
|---|---|---|---|
| Budget Alerts | Spend reaches a percentage of a budget you set | Every 10 minutes | Live |
| Daily Reports | Your daily spend & usage digest is due | Every 5 minutes | Live |
| Hanging Requests | Any single request runs longer than your threshold | Every 5 minutes | Live |
| LLM Too Slow | p95 latency across the last 10 minutes exceeds your threshold | Every 5 minutes | Live |
| LLM Exceptions | The error rate across the last 10 minutes reaches your threshold | Every 5 minutes | Live |
| Outage Alerts | One provider fails at least your threshold share of requests | Every 5 minutes | Live |
| Region Outage | One regional endpoint fails at least your threshold share of requests | Not checked | Preview |
Region Outage rules are saved and kept exactly as you enter them, but nothing evaluates them yet, so no notification is sent even when the condition is met. The Alerts page labels that rule Not monitored yet so you never have to infer it from silence.
When an alert actually fires
Saving a rule is not the same as meeting its conditions. Four things decide whether you get a notification, and each of them has caught somebody out.
1. A channel must be bound
Telemetry alerts (Hanging Requests, LLM Too Slow, LLM Exceptions, Outage, Region Outage) deliver only to the channels you tick under Deliver to channels. A rule with no channel bound evaluates and notifies nobody — the Alerts page flags it as No channel — cannot notify.
Budget Alerts and Daily Reports are the exception: they email your owners and admins with no channel setup at all.
2. The window needs enough traffic
The three aggregate alerts evaluated today are measured over a rolling 10-minute window, and a window with fewer than 20 measurable requests produces no verdict at all — not a pass, not a failure. A percentile or a rate over three requests is noise, and paging on noise is worse than staying quiet.
| Alert | Needs, inside one 10-minute window |
|---|---|
| LLM Too Slow | 20 requests with a recorded duration |
| LLM Exceptions | 20 requests total |
| Outage Alerts | 20 requests to that one provider |
Region Outage is not in this table because it is not evaluated at all yet — its silence has a different cause, and no amount of traffic changes it.
This is the most common reason a correctly configured alert never fires. At roughly 15 requests a day, no 10-minute window will ever hold 20 of them, so a threshold you can clearly see being exceeded in your logs still produces no alert. Hanging Requests has no such floor — it compares each request on its own, so one qualifying request is enough.
That makes Hanging Requests the only alert that reaches low-volume traffic, and it covers latency only. There is no per-request error alert: to catch failures below the 20-request floor, watch them in Analytics & Reports rather than waiting on a notification.
3. Requests need a moment to settle
A request that finished in the last 30 seconds is not counted yet. Without that margin an in-flight request looks identical to a stalled one, and every alert would open with a burst of false pages. It means a genuine problem is reported on the next check rather than the current one.
Hanging Requests looks back over the last 60 minutes for individual slow requests; the aggregate alerts look at the last 10 minutes.
(None of this applies to Region Outage — it is not evaluated yet, so it sends nothing regardless of traffic or threshold.)
4. One page per problem, not one per check
Each rule holds a 15-minute cooldown after it notifies, so a condition lasting an hour reaches you a handful of times instead of once every check. Budget Alerts use a 24-hour cooldown for the same reason — a low balance is still low five minutes later, and that is not news.
Worked example
You set Hanging Requests to 2 seconds and bind an email channel. At 10:00:04 a request takes 6 seconds.
- 10:05 — the check runs. The request finished at 10:00:10, more than 30 seconds ago, so it counts. 6 s is over your 2 s threshold, so you are emailed, and the message names the offending requests and their models.
- 10:10 and 10:15 — more slow requests, but the rule is inside its 15-minute cooldown, so no second email.
- 10:20 — the cooldown has passed. If requests are still hanging, you are emailed again.
Set the same 2 seconds on LLM Too Slow instead and you may hear nothing at all: unless 20 requests land inside one 10-minute window, there is no p95 to compare.
Delivery
Telemetry alerts (LLM Exceptions, LLM Too Slow, Hanging Requests, Outage Alerts) are evaluated on a schedule and delivered to every channel bound on the rule. Set the threshold and tick the channels on the Alerts page; channels themselves are created under Alert Channels — see Observability & Logs. Region Outage accepts the same configuration but is not evaluated yet.
Credit balance alerts are emailed to your organization's owners and admins.
Daily Reports are an email digest: switch the toggle on and a daily spend & usage summary is emailed to your organization's owners and admins. There is no channel or threshold to configure — the toggle is the whole setup, and switching it off stops the digest.
Pausing an alert
Need quiet during a known incident or maintenance window? An owner or admin switches the alert off with its toggle before the window and back on afterward. There is no separate snooze timer — the toggle is the pause.
Next steps
- Observability & Logs — Define the channels alerts deliver to
- Budget Controls — Set the budgets that Budget Alerts watch
- Analytics & Reports — Investigate what triggered an alert
FAQ
Who can create or change alerts?
Only organization owners and admins can enable, disable, or configure alerts on the Alerts page. Members and viewers see the current settings in read-only mode and need to ask an admin to make changes.
Which alerts can I actually turn on today?
All of them can be turned on, and all but one are evaluated and delivered: Budget Alerts, Daily Reports, Hanging Requests, LLM Too Slow, LLM Exceptions and Outage Alerts. Region Outage is the exception — the rule is saved and kept exactly as you set it, but nothing evaluates it yet, so it sends nothing. The Alerts page marks that rule Not monitored yet.
I can see the threshold being exceeded in my logs, so why no alert?
Almost always the 20-request floor. LLM Too Slow, LLM Exceptions and Outage Alerts are measured over a rolling 10-minute window and produce no verdict at all below 20 measurable requests in that window — a p95 or an error rate over a handful of requests is noise. Low-volume traffic never reaches the floor, so the rule stays silent no matter how far past the threshold an individual request goes.
Three other things to check, in order: the rule has a channel bound (telemetry alerts deliver only to bound channels); the request finished more than 30 seconds ago (anything newer has not settled yet); and the rule is not inside its 15-minute cooldown from a previous notification.
For low-volume traffic, use Hanging Requests — it compares each request on its own with no sample floor, so a single slow request pages you.
How quickly will I hear about a problem?
Telemetry alerts are checked every 5 minutes and Budget Alerts every 10, so allow one check interval plus the 30-second settle margin. Worst case for a telemetry alert is a little over 5 minutes from the request finishing to the notification landing.
Why did I only get one email when the problem lasted an hour?
Each rule holds a 15-minute cooldown after it notifies (24 hours for Budget Alerts). A sustained problem reaches you several times rather than on every single check.
How do Daily Reports arrive, and can I change the recipients?
The daily digest is emailed to your organization's current owners and admins — the recipient list follows your membership automatically, so promoting or removing an admin updates who receives it. It can't be routed to a Slack/Teams/webhook channel.
Do alerts stop or throttle my requests?
No. Alerts are notification-only — they tell you something happened but never block, pause, or slow traffic. The hard stop when a budget or credit balance is exhausted is enforced separately by Budget Controls, which returns a 402 at the limit.
What notification channels can I use?
Email, Slack, Microsoft Teams, Jira, and a generic Webhook. Channels are created and managed under Alert Channels (see Observability & Logs); PagerDuty is reachable through a generic webhook pointed at its Events API.
Do alerts consume my credits or count against my rate limits?
No. Alerts don't make model calls, so they don't draw down your credits or count toward your RPM/TPM limits — they're a platform notification feature that sits alongside your usage.
What threshold controls a budget alert?
Currently, a soft budget only returns an x-nr-budget-warning header when exceeded, and does not dispatch any notifications.
Is there a history log of alerts that fired?
No. The Alerts page shows your current rules and their delivery status, but it doesn't keep a log of individual fired alerts. To investigate what triggered one, use your request logs and Analytics & Reports.
How do I pause an alert during planned maintenance?
An owner or admin can switch a live alert off with its toggle before the maintenance window and turn it back on afterward, so no notifications fire while you're expecting disruption.
Observability & Logs
Gain full observability into LLM traffic with nRouter. Inspect request logs, analyze latency, configure PII redaction, and stream logs to external SIEMs.
Router Settings
Configure intelligent routing rules, model weights, automated failovers, and caching policies in nRouter to optimize latency, cost, and reliability goals.