Observability Overview

The NVIDIA NeMo Guardrails library exposes three observability signals so you can debug locally during development and monitor behavior in production: logs, traces, and metrics. Each signal targets a different question and can be enabled independently of the others.

Signal	Best For	Page
Logging	Debugging a single request with verbose console output, the explain() method, and the log generation option for structured per-request data.	Logging and Debugging Guardrails Generated Responses
Tracing	Following a request through the rails it activated and the LLM calls it issued, with full OpenTelemetry semantic-convention support.	Tracing Guardrails
Metrics	Tracking aggregate behavior, including request volume, latency distributions, error rates, saturation, and per-LLM-call token usage for SLO dashboards and alerting.	Metrics for Guardrails

Tracing and metrics use the OpenTelemetry library-instrumentation pattern: the NVIDIA NeMo Guardrails library depends on the OpenTelemetry API only, and the host application configures the SDK providers, exporters, and processors.