Skip to Content

How do I move my alerts from Traceloop to Rius?

Rius covers the metric and error side of Traceloop monitoring with four pre-defined alerts and custom alert rules, which you set up once OpenLLMetry sends traces to Rius. Your OpenLLMetry instrumentation stays as it is and only its export settings change. Rius alerts do not judge output quality, so Traceloop evaluators and LLM-as-a-judge monitors have no counterpart. Rius sends alerts to webhook, email and Slack channels.

How it works

Rius alerts run on the traces you send, so the move has two parts: get your spans into Rius, then recreate each alert you rely on. Rius starts from the first span you send it, so alerts evaluate new traffic only.

Here is how the Traceloop concepts map to Rius.

  • Metric and error monitors map to custom alerts. A custom rule can watch cost, input, output and reasoning tokens, duration, time to first token, context usage, error rate, LLM error rate, errors, traces and LLM calls. You can filter by service, agent, step, model, span kind, operation, session or instance.
  • Looping or failing agents map to the four pre-defined alerts: Consistent Error Pattern, Swallowed Error, Tool Loop and Context rot. You can turn them off but not delete them.
  • Notification targets map to Rius alert channels: webhook, email and Slack.
  • Evaluators and LLM-as-a-judge monitors have no counterpart. Rius has no evaluators running on production spans, so nothing in Rius scores answer quality. Rius alerts cover errors, loops and metric thresholds.

Your decorators carry over. OpenLLMetry’s @tool span arrives in Rius as a TOOL span and @agent as an AGENT span, which matters for the Tool Loop alert. That alert counts only spans of kind TOOL, so a tool traced as a generic step is never counted. Rius names the agent after your service.name, which OpenLLMetry takes from app_name or appName.

Set it up

  1. Point OpenLLMetry at Rius. Change the base URL and the auth header, and in Python turn off metrics, because Rius ingests traces only. If your Traceloop monitors must keep firing while you rebuild them, send OpenLLMetry to a Collector that exports to both backends instead, as the side-by-side setup shows. Migrate from Traceloop covers both paths, including the TypeScript settings.

    # After: Rius export TRACELOOP_BASE_URL="https://ingest.eu.console.rius-glassflow.com" export TRACELOOP_HEADERS="Authorization=Bearer%20<your Rius API key>" export TRACELOOP_METRICS_ENABLED="false" # Python SDK only
  2. List the monitors you use today. For each one, note what it measures and where it notifies. Sort them into threshold or error monitors, which move to Rius, and evaluator monitors, which do not.

  3. Check the pre-defined alerts first. Consistent Error Pattern, Swallowed Error, Tool Loop and Context rot are already available, so a loop or repeated-failure monitor may need no new rule. Swallowed Error counts errored child steps that are not tool calls.

  4. Recreate threshold monitors as custom alerts. Pick the measure, the threshold and the filters. A rule measures only the spans it matches, and tokens and cost sit on model-call spans (LLM, embedding, reranker). A filter that selects only tool, chain or agent spans therefore measures no tokens and no cost. To alert on model spend, filter by Model, or set Span kind to LLM, and add embedding and reranker with in if your agent calls them. To measure whole runs, filter by Service or Agent.

  5. Add your channels. Create the webhook, email or Slack channel that each alert should notify.

  6. Compare, then retire. With the Collector fan-out from step 1, both backends see the same traces, so compare what fires on each side, then retire the Traceloop monitors. The cutover checklist lists moving monitors to alerts as a step of its own.

What you get

  • Alerts on errors, loops, cost, tokens, duration, time to first token, context usage and call counts, delivered to webhook, email or Slack.
  • Four pre-defined alerts that need no setup, including Tool Loop for agents that repeat the same tool call.
  • Your existing OpenLLMetry decorators and traces, unchanged in code, with cost computed by Rius from the model and token counts. Rius costs are list-price estimates in US dollars.
  • A clear boundary: alerts watch how your agents run, not how good their answers are.

When something else fits

  • If an alert depends on an evaluator, LLM-as-a-judge or guardrail, keep that workflow in Traceloop and send traces to both backends through the Collector fan-out. Rius has no evaluators and nothing that blocks or rewrites model output.
  • In Python, re-pointing OpenLLMetry at Rius switches off Traceloop platform features such as get_prompt(), guardrails and datasets. If your code uses them, use the fan-out instead of re-pointing, as Migrate from Traceloop explains.

Still stuck? Get help

Last updated on