A company is building a public-facing chatbot that must not generate offensive content. They plan to use a combination of guardrails. Which combination best represents a layered defense?

Input guardrail that filters offensive prompts, and output guardrail that filters offensive responses.

A system prompt that instructs the model to be polite.

Only an input guardrail that filters offensive prompts.

Only an output guardrail that filters offensive responses.

A team has implemented both input and output guardrails for their customer-support chatbot. After deployment, they observe that the input guardrail blocks too many legitimate queries, while the output guardrail occasionally misses harmful content. They need to improve the guardrails without sacrificing user experience. Which approach is most effective?

Collect data on blocked queries and missed harmful outputs, then use this data to tune the guardrail thresholds and rules.

Switch to a more powerful LLM that is less likely to generate harmful content.

Increase the strictness of both guardrails to ensure no harmful content gets through.

Remove the input guardrail and rely solely on the output guardrail to catch harmful content.

Which of the following is an output guardrail?

Limiting the number of tokens in the prompt

Filtering the user's prompt to remove offensive words

Adding instructions to the prompt to avoid harmful content

Checking the response for factual accuracy against a knowledge base

A team is conducting a qualitative evaluation of an LLM's responses to customer complaints. They notice that some responses are polite but do not actually address the customer's issue. Which type of issue is this, and how should it be categorized?

Quality issue: factual inaccuracy.

Which of the following is a quality issue in an LLM response?

The response is off-topic and does not answer the user's question

The response contains a personal phone number

The response includes hate speech

The response provides instructions for illegal activity

DatabricksCertified Generative AI Engineer Associate

Domain 3Objective 4

Safety and Guardrails GENERATIVE-AI-ENGINEER-ASSOCIATE Practice Questions (Page 6)

Part of the Section 3: Application Development domain, which makes up ~18% of our current practice bank. Databricks does not publish an official question count, but from its 90-minute exam (~35–60 total, ~6–11 in this domain), expect 1–2 from this objective — we provide 31 practice questions to prepare you well beyond it. (estimate)

31questions here

7free pages

7concepts

Questions 26–30

26application · medium
A company is building a public-facing chatbot that must not generate offensive content. They plan to use a combination of guardrails. Which combination best represents a layered defense?
Select an answer first
27expert · hard
A team has implemented both input and output guardrails for their customer-support chatbot. After deployment, they observe that the input guardrail blocks too many legitimate queries, while the output guardrail occasionally misses harmful content. They need to improve the guardrails without sacrificing user experience. Which approach is most effective?
Select an answer first
28foundation · easy
Which of the following is an output guardrail?
Select an answer first
29application · medium
A team is conducting a qualitative evaluation of an LLM's responses to customer complaints. They notice that some responses are polite but do not actually address the customer's issue. Which type of issue is this, and how should it be categorized?
Select an answer first
30foundation · easy
Which of the following is a quality issue in an LLM response?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to GENERATIVE-AI-ENGINEER-ASSOCIATE

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by Databricks. “GENERATIVE-AI-ENGINEER-ASSOCIATE” is a trademark of its owner, used for identification only.