AIF-C01 Practice Question: An e-commerce company is building a product…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Amazon Comprehend for toxicity detection

  Why it's wrong here

  Comprehend toxicity detection classifies existing text after generation, adding latency and cost while leaving competitor mentions undetected, since it scores toxicity rather than custom denied terms. It suits post-hoc analysis of stored content, whereas Bedrock Guardrails blocks prohibited topics and keywords during generation.

• ✓

  Bedrock Guardrails with content filters and denied topics

  Why this is correct

  Guardrails content filters screen prompts and responses for harmful categories, while denied topics block the specific prohibited subjects and competitor mentions the company listed. This directly enforces the stated content restrictions at inference time without retraining or prompt engineering.

• ✗

  Bedrock Knowledge Bases with metadata filtering

  Why it's wrong here

  Metadata filtering restricts retrieval to documents matching attribute tags; it cannot inspect generated output or block denied topics and keywords. Knowledge Bases ground responses in your own corpus, so they suit retrieval-augmented answers over proprietary data, not content moderation of model completions.

• ✗

  Bedrock Agents with a custom action group to filter outputs

  Why it's wrong here

  Bedrock Agents orchestrate multi-step tasks via action groups, but they lack native content filtering logic; filtering would require custom code to invoke a separate moderation API, adding latency and complexity. This option tempts because action groups can call external systems, so one might think to build a filter there. However, the correct choice—Bedrock Guardrails—provides built-in, low-latency keyword and topic blocking without custom development.

Go deeper

Related to this question