Implement generative AI guardrails with Azure AI Content Safety: Review

The groundedness detection exercise flagged a sample AI response as 'ungrounded' because it mentioned 'automatic scaling'—a feature not in the source text. Cool concept. I wanted to try my own text, maybe something more subtle. But the pre-written notebook was locked down. I couldn't change the input. I couldn't even see the system prompt that led to the original answer. You just click 'Run' on a cell, watch a spinner, and read the JSON output. It’s a guided tour, not a lab.

This pattern repeats for text moderation, image moderation, and the new prompt shields. The module does a fantastic job showing you what the API endpoints return for pre-canned examples of hate speech, self-harm, and jailbreak attempts. The severity levels (0, 2, 4, 6) are clearly explained. You get a perfect, sterile demonstration of the service. But I don't feel like I *learned* how to implement guardrails. I learned how to call an API and parse its output. There's no discussion of what to do with a 'severity: 4' flag, how to chain these checks, or how to handle edge cases. It’s a product catalog page made interactive. It’s useful for exactly that: seeing what Azure offers before you commit to reading the real, much longer, documentation. It gave me the right keywords to search for later.