AI Intermediate: AWS Bedrock & AI Services Practice Questions

Amazon Bedrock Guardrails is the purpose-built feature that covers all three requirements in a single, integrated configuration. Denied topics can be configured to block references to specific competitor brand names. Content filters with configurable thresholds can block sexually explicit material. PII redaction can be enabled to automatically detect and redact entities such as email addresses from prompts before they reach the model. Model Evaluation with a custom rubric is incorrect because Model Evaluation assesses model output quality; it does not enforce runtime content policies or redaction. A pre-processing Lambda in an action group is incorrect because, although it could handle preprocessing, it is not the recommended native Bedrock approach and would need custom code to replicate what Guardrails provides natively; it also does not filter content on the model output side. Macie combined with Comprehend is incorrect because, while these are valid AWS services, combining them externally adds unnecessary complexity and latency, and this approach still does not apply guardrails to the model's outputs. Bedrock Guardrails handles both input and output.

HTTP 429 means 'Too Many Requests' (throttling). Amazon Bedrock returns it when a caller exceeds the requests-per-minute or tokens-per-minute quota for a model. The recommended fix is to wait using exponential backoff with jitter, then retry. The missing IAM permission option is wrong because a permission error returns HTTP 403 (AccessDeniedException). The Region availability option is wrong because that problem returns a 4xx validation error or ResourceNotFoundException, not 429. The context window option is wrong because an oversized prompt produces a ValidationException (400), not 429.

Bedrock Agents require action group Lambda functions to return a specific response structure. If the Lambda returns an empty or malformed body, the agent may treat it as a successful operation with no data and keep reasoning, which leads to fabricated answers. The missing lambda:InvokeFunction permission option is wrong because a permission problem causes an explicit authorization error that the agent would surface, not a silent drop. The missing 4xx error schema option is wrong because OpenAPI error definitions help the agent understand API contracts, but the root issue is the Lambda response format. The missing instruction guidance option is wrong because instructions can shape error handling, but the underlying problem is that the Lambda never returns a proper response that signals failure.

Provisioned throughput reserves model units (MUs) for exclusive use by the customer and is billed on an hourly basis for the committed term, regardless of whether tokens are actually consumed — making the reserved-capacity, hourly-billing statement correct. Provisioned throughput is economically beneficial only when utilization is consistently high enough to justify the committed cost; at low or variable utilization, on-demand is cheaper — making the high-utilization statement correct. On-demand is NOT always cheaper; at sustained high throughput, provisioned throughput can be significantly more cost-effective. Provisioned throughput is not limited to Titan models; it supports many Bedrock models including Claude. While provisioned throughput is required to host fine-tuned custom models, this is an additional feature distinction, not the primary pricing comparison point — and the custom-model statement as written is not accurate as a blanket rule. On-demand does scale automatically, which is true but was not among the correct answer choices selected.

Bedrock Guardrails includes a sensitive information filter that can detect and then redact or block specific PII entity types, including Social Security Numbers and credit card numbers, in both input prompts and model responses. Enabling this filter with the relevant entity types directly addresses the compliance concern. A denied topic for 'financial personal data' blocks responses on a defined subject area, but topic-level blocking is too coarse to reliably prevent PII. It would not reliably catch SSNs embedded in otherwise compliant responses. Content filters for sexual content and violence have nothing to do with PII. Those categories cover harmful content types, not personal data exposure. Contextual grounding checks keep responses anchored to source documents. They do not detect or redact PII data types, and they would not stop the model from surfacing PII that appears in retrieved content.