AWS Certified Generative AI Developer – Professional 1-10

This question has three requirements: structured inputs, a consistent output format, and detecting and notifying on bullying language without blocking it.
With A, Amazon Bedrock Prompt Management lets you define the four required inputs as variables and specify the output format, so it delivers both structured inputs and a consistent response format.
The key point of D is that setting the insults filter’s response action to “detect” rather than “block” lets you flag and notify while still returning the response.
B uses the block setting, which violates the requirement not to block everything.
The prompt router in C is unnecessary for a single-model configuration, and the inference profile in E does not guarantee enforcement of variables or the output format.
Detect harmful content with content filters – Amazon Bedrock

For a workflow that inserts human approval, the AWS Step Functions callback pattern is ideal.
Using waitForTaskToken lets you pause execution until a human completes the review and calls SendTaskSuccess, reliably returning the approval result to the workflow.
Storing the review decisions in Amazon DynamoDB also provides the durable record needed for auditing.
A relies on Glue, whose visual workflows are intended for ETL and are not suited to waiting on human approval.
C uses ElastiCache to cache decisions, which is unsuitable for the durable storage required for auditing.
D is a custom implementation that makes state management and timeout control complex, imposing a higher operational burden than managed Step Functions.
Wait for a callback with the task token – AWS Step Functions

The requirements are threefold: proactive alerting as limits are approached, scaling beyond 5,000 requests per minute, and maintaining metrics for cost allocation.
Estimating token counts before sending, publishing metrics to CloudWatch, and raising alarms at a threshold lets you warn proactively before the limit is reached, and recording data in DynamoDB enables per-unit cost allocation.
A analyzes error logs after the fact, so it does not provide proactive alerts.
The API Gateway rejection in C is throttling and does not meet the requirements for proactive notification or cost allocation metrics.
Bedrock Guardrails in D is a content-safety feature and is not intended for token quota management.
Monitor Amazon Bedrock with Amazon CloudWatch

The requirement is comprehensive observability that consolidates both technical and business metrics in one place and automatically alerts on degradation.
Bringing both sets of metrics into a custom CloudWatch dashboard, catching correlated degradation with a composite alarm that uses anomaly detection, and notifying via SNS is the configuration that best matches the requirements.
C is close but lacks anomaly detection and correlation analysis, remaining a simple threshold-based approach.
A goes beyond the “provide alerts” requirement by performing automatic remediation with Grafana, which is excessive.
B streams to QuickSight via S3, which has low real-time performance and is inferior for the immediacy of operational alerts.
Create a composite alarm – Amazon CloudWatch

The requirements are fourfold: resilience to transient errors, prevention of cascading failures, distributed tracing across service boundaries, and correlation with FM characteristics.
The AWS SDK’s standard retry mode provides exponential backoff with jitter, which avoids concentrated simultaneous retries and prevents cascading failures; and adding annotations such as the model ID in AWS X-Ray lets you analyze latency factors by FM characteristic.
The fixed-delay retry in A tends to synchronize retries and does not provide distributed tracing.
C uses only caching and logging, which cannot prevent cascading failures and offers no cross-service tracing.
CloudTrail in D is for audit logging and is not distributed tracing intended for request latency analysis.
Retry behavior – AWS SDKs and Tools Reference Guide

The crux of this question is giving different visibility of the same data by role.
The basis for the correct answer is that, rather than removing PII at ingestion, you dynamically apply a guardrail configuration based on the Cognito group at inference time, returning full responses to surgeons and masked responses to engineers.
An S3 lifecycle automatically deletes documents older than three years, and the daily sync keeps the knowledge base current.
A permanently strips PII at ingestion, which violates the surgeon’s viewing requirement.
B fixes masking at ingestion and cannot provide role-based control.
D duplicates the knowledge base, increasing cost and the risk of data inconsistency.
Amazon Bedrock Guardrails

The requirement is to block inappropriate content, prevent PII exposure, and integrate with the S3 workflow, all while minimizing infrastructure management.
Combining managed Bedrock guardrails, Comprehend PII detection, and Rekognition image moderation with Step Functions avoids building custom models and minimizes operational burden.
A has many redundant monitoring alarms and does not directly enforce content safety.
B requires building and maintaining a custom model with SageMaker, which is a heavy burden.
C performs moderation at authentication time, and Textract is not intended for moderation, so it is inappropriate.
Moderating content – Amazon Rekognition

To enforce policy strongly while reducing false positives, it is recommended to set different actions per content type within a single guardrail.
Setting the filter to “Medium” rather than “High” avoids over-blocking legitimate customer interactions while still preventing harmful content, and you can use different handling by context—masking PII in responses and blocking financial information in inputs.
Adding a definition and sample phrases to the denied topic improves detection accuracy.
A blocks all PII uniformly, which causes frequent false positives.
B is unnecessarily complex and is not how guardrails are meant to be used.
D chains guardrails with Step Functions redundantly, which guardrails can natively handle.
Components of a guardrail – Amazon Bedrock

The root cause of the problem is that the search scope is too broad.
Amazon Bedrock Knowledge Bases supports metadata-aware filtering, and by indexing S3 object metadata (such as content type, publication date, and source) to narrow the search scope, you can reduce the number of documents evaluated and improve both latency and relevance.
This is the approach that requires the least architectural change.
A is re-training the embeddings, which has already been judged ineffective.
Migrating to OpenSearch in B and to Amazon Q Business in D both increase operational burden and implementation effort and deviate from the existing configuration.
Amazon Bedrock Knowledge Bases

Amazon Bedrock Prompt Management is a purpose-built feature for governing and standardizing prompts at scale across teams and Regions.
The basis for the correct answer is that it provides native version control and approval workflows, lets you control approval permissions with IAM policies, and integrates with CloudTrail to leave an immutable audit trail.
Parameterized templates let you enforce consistent quality standards.
A requires substantial custom development to coordinate approvals.
C relies on general-purpose services and manual version control, which is error-prone as it scales.
D uses a service combination not designed for governing generative AI prompts at scale and is unnecessarily complex.
Prompt management in Amazon Bedrock