groups: - name: EmbeddedExporter rules: # The threshold (1) is in USD. The `model` label carries the resolved model-name (post-routing). # PromQL `increase()` requires ≥2 datapoints with growth-difference to extrapolate positive — # for brand-new counter series this needs ≥2 distinct request bursts ≥1 scrape-cycle apart. - alert: LitellmProviderSpendOverBudget expr: 'sum(increase(litellm_spend_metric_total{model=~"(claude-|anthropic/).*"}[24h])) > 1' for: 5m labels: severity: warning annotations: summary: LiteLLM provider spend over budget (instance {{ $labels.instance }}) description: "Cumulative spend for an LLM provider has exceeded the daily budget threshold. Replace the regex `(claude-|anthropic/).*` with your provider's model-name pattern. Useful as a soft-warning when `provider_budget_config` hard-cap is unavailable or disabled.\n VALUE = {{ $value }}\n LABELS = {{ $labels }}" - alert: LitellmProxyFailedRequestsRateHigh expr: 'sum(rate(litellm_proxy_failed_requests_metric_total[5m])) / sum(rate(litellm_proxy_total_requests_metric_total[5m])) > 0.05' for: 10m labels: severity: warning annotations: summary: LiteLLM proxy failed requests rate high (instance {{ $labels.instance }}) description: "LiteLLM proxy is returning failed responses to clients (>5% error rate over 5min). Investigate downstream LLM provider availability or auth issues.\n VALUE = {{ $value }}\n LABELS = {{ $labels }}" - alert: LitellmRequestLatencyP95High expr: 'histogram_quantile(0.95, sum(rate(litellm_request_total_latency_metric_bucket[5m])) by (le)) > 10' for: 10m labels: severity: warning annotations: summary: LiteLLM request latency p95 high (instance {{ $labels.instance }}) description: "LiteLLM request total latency p95 exceeds 10 seconds over 5min. Check downstream LLM provider response-times and proxy queue-depth.\n VALUE = {{ $value }}\n LABELS = {{ $labels }}"