Skip to lesson

NVIDIANVIDIA-Certified Professional: Agentic AI2026-109.2 Integrate compliance guardrails (privacy, enterprise policy).

Topic Control for an IP-Law Research Assistant

Completed by opening the official lab

Why it matters

A system-prompt instruction to stay on topic is advisory. A separate classifier enforces the policy before the main model sees the prompt, and how you write that policy decides what it blocks.

What to practise

Configure a type: topic_control model on engine: nim with llama-3.1-nemoguard-8b-topic-control, add the topic safety check input $model=topic_control flow, and write the IP-law policy as the prompt for that task. Run the smoke test and the evaluation, then read recall per category. Pay attention to adjacent legal questions, the hardest case, and trace each failure to its cause: a compound-domain prompt or a term the policy never mentions.

How it fits

This is the topic-control half of 9.2.2. Retention settings, the other half, sit outside the rail and need their own owner. Compare the false positive here with the overblocking that 9.3.2 measures.