Databricks · Databricks Generative AI Engineer Associate
Quality and Guardrails
Databricks Generative AI Engineer Associate · Lesson 8 of 17 · 55:15
YouTube loads only after you press play. This chapter opens at its timestamp in the full course. Watch on YouTube ↗
In this lesson
Prepare for the Databricks Certified Generative AI Engineer Associate exam. In this lesson: common problems (hallucination, off-topic, incomplete, wrong format or tone), safety risks (toxic content, leaked data, prompt injection), the MLflow judges that spot them and the fix for each, Unity Gateway service policies as guardrails (AI Gateway is legacy), and approved endpoints with Can Query only. Includes three original practice questions.
Preview the topics
- A customer asks about the warranty on the CX-450, and the assistant confidently answers two years. The retrieved manual pages say nothing about the length of the warranty. Which problem is this?
- Copperline wants to stop card numbers in customer messages from reaching any model, and to catch attempts to override the system prompt. The rules must apply to every model, governed in Unity Catalog. What should the team set up?
- Copperline turned on guardrails for its approved serving endpoint. Then a developer created a new endpoint, without any guardrails, and pointed the assistant at it. What would have prevented this?