Tracing Guardrails Quickstart

Why it matters

An audit can rely only on what the system recorded while it ran. This notebook shows which guardrail decisions NeMo Guardrails writes down for each request, and in what order.

What to practise

Build two configurations that differ only in rails.input.parallel. Each has content-safety, topic-control and jailbreak-detection input rails and a content-safety output rail. Enable tracing with the FileSystem adapter and a JSONL filepath. Send one safe and one unsafe request through LLMRails in each mode. Then plot each rail’s duration and a Gantt chart from the spans. Check that both modes return the same responses, that input-rail time is the sum of the three rails in sequential mode and the longest one in parallel mode, and that the unsafe request’s trace shows neither the application LLM nor the output rail.

How it fits

The trace is the recorded execution that 10.3.2 separates from a generated explanation. Compare the FileSystem adapter with the OpenTelemetry route that 10.3.2 describes, and decide which one your decision records should depend on.