Select Two (Multi-Select) Questions

Configure Macie to automatically grant access to the SageMaker execution role

Why wrong: Macie is for data discovery, not access control.

Create a VPC endpoint for S3 and associate it with the SageMaker VPC

Why wrong: VPC endpoints are not required for cross-account access; they are used for private connectivity.

Remove the bucket policy that prohibits external access

Why wrong: That would weaken security; the policy can be modified to allow specific cross-account access.

Amazon RDS for MySQL

Why wrong: RDS is not supported as a direct knowledge base source.

Amazon DynamoDB

Why wrong: DynamoDB is not supported as a direct knowledge base source.

AWS Glue

Why wrong: Glue is an ETL service, not a data source for knowledge bases.

Enable multi-factor authentication (MFA) for all AWS accounts.

Why wrong: MFA provides an extra layer of login security but does not prevent authorized users from deploying.

Use AWS CloudTrail to audit deployment actions.

Why wrong: CloudTrail logs actions but does not enforce restrictions.

Use Amazon GuardDuty to monitor for unauthorized deployment.

Why wrong: GuardDuty detects threats but does not prevent authorized users from deploying.

Enable Amazon CloudWatch Logs to capture model output and manually review.

Why wrong: Manual review is not scalable and does not enforce guidelines in real time.

Use AWS IAM policies to restrict the model's output to only approved words.

Why wrong: IAM policies cannot restrict the content of model responses.

Encrypt the model responses using AWS KMS to prevent unauthorized viewing.

Why wrong: Encryption protects confidentiality, not content compliance.

AWS CloudTrail

Why wrong: CloudTrail logs API calls but does not govern model lifecycle.

Amazon S3

Why wrong: S3 stores artifacts but does not provide lifecycle governance.

AWS CodePipeline

Why wrong: CodePipeline is for CI/CD, not model governance.

Amazon SageMaker Model Monitor for drift detection

Why wrong: Model Monitor detects data and model quality drift, not content moderation.

Amazon Rekognition for image moderation

Why wrong: Rekognition is for image and video moderation, not text content.

Amazon Comprehend for sentiment analysis

Why wrong: Sentiment analysis helps detect bias but does not block harmful content in real-time.

Always use the largest available model for best results

Why wrong: Using the largest model increases cost without guaranteed throughput benefit.

Fine-tune a large foundation model on the company's codebase

Why wrong: Fine-tuning incurs additional training and hosting costs, not a cost-saving measure.

Apply an IAM policy that denies access to the conversation logs for the user who requested deletion.

Why wrong: Denying access does not delete the data, which is required by GDPR.

Disable encryption of conversation logs to allow manual deletion.

Why wrong: Encryption is separate from deletion; disabling does not help.

Modify the Lex bot's conversation log settings to stop collecting personal data.

Why wrong: Stopping collection does not delete already collected data.

Principal Component Analysis (PCA)

Why wrong: PCA is an unsupervised dimensionality reduction algorithm.

Latent Dirichlet Allocation (LDA)

Why wrong: LDA is an unsupervised topic modeling algorithm.

K-Means

Why wrong: K-Means is an unsupervised clustering algorithm.

It can only be used for text data.

Why wrong: It supports image, video, and text.

It can only use a public workforce from Amazon Mechanical Turk.

Why wrong: It also supports private workforces (Vendors and your own employees).

Set missing values to zero

Why wrong: Setting to zero may introduce bias and assumes the feature value is zero, which is often incorrect.

Ignore missing values during training

Why wrong: Most algorithms cannot handle missing values natively and will throw errors or produce incorrect results.

Replace missing values with -1

Why wrong: Using -1 assumes the feature can be negative; it may not be appropriate and can distort distances.

Increase the number of epochs

Why wrong: More epochs often lead to overfitting.

Use a smaller learning rate

Why wrong: Smaller learning rate may not prevent overfitting directly.

Combine multiple models to obscure decision logic

Why wrong: Obscuring logic is opposite of transparency.

Remove all features except the most predictive ones

Why wrong: Removing features may reduce explainability.

Use black-box models to protect proprietary algorithms

Why wrong: Black-box models reduce transparency.

Amazon Redshift

Why wrong: Redshift is a data warehouse, not a vector store.

Amazon RDS for MySQL

Why wrong: RDS for MySQL does not natively support vector operations.

Amazon DynamoDB

Why wrong: DynamoDB is a NoSQL database, not a vector store.

Amazon SageMaker

Why wrong: SageMaker is for model training and deployment, not a core RAG component.

AWS Lambda

Why wrong: Lambda can be used in action groups for agents, but not as a primary RAG component.

Amazon Bedrock Guardrails

Why wrong: Guardrails are optional safety features, not part of the core RAG pipeline.

A vector store like Amazon OpenSearch Serverless

Why wrong: Vector stores are used for semantic search, not for executing API calls or database updates.

Bedrock Guardrails

Why wrong: Guardrails handle content filtering and safety, not external system integration.

Bedrock Knowledge Bases

Why wrong: Knowledge Bases are for RAG, not for executing external actions.

A fine-tuned foundation model specific to the support domain

Why wrong: Fine-tuning is optional; the agent can use a base model with RAG.

A custom model trained on historical support tickets

Why wrong: Custom models are not required; pre-trained models are sufficient.

Reduce the temperature parameter to 0

Why wrong: Lower temperature reduces randomness but does not prevent hallucinations; the model may still fabricate facts.

Fine-tune the LLM on a general dataset of medical texts

Why wrong: Fine-tuning may not eliminate hallucinations; the model can still generate plausible but incorrect information.

Use Bedrock Guardrails with a content filter for medical topics

Why wrong: Content filters block certain categories of content but do not ensure factual accuracy; they may block appropriate medical advice.

Automatically retrain the model weekly

Why wrong: Retraining is an operational practice, not documentation.

Set up Amazon CloudWatch alarms to monitor model accuracy

Why wrong: CloudWatch alarms monitor metrics but do not provide documentation of model limitations.

Conduct A/B testing between different model versions

Why wrong: A/B testing compares model performance, not documentation.

Remove protected attributes (e.g., gender, race) from the dataset

Why wrong: Removing attributes may not eliminate bias because other correlated features can act as proxies.

Adjust the decision threshold to increase approval rates for the disadvantaged group

Why wrong: Adjusting only the threshold may mask bias without addressing its root cause; it can also hurt model performance.