Foundation Model Integration Data And Compliance practice questions

Store all data in-memory only to avoid persistent storage risks.

Why wrong: This does not provide durability or meet the auditability requirements of HIPAA.

Use public S3 buckets with ACLs enabled.

Why wrong: Public buckets are a security risk and violate HIPAA compliance.

Disable logging to prevent sensitive data from appearing in logs.

Why wrong: Auditing and logging are mandatory for HIPAA compliance.

The model's training date.

Why wrong: The training date is less important than the ability to follow retrieval-augmented instructions.

The model's ability to generate images.

Why wrong: Image generation is irrelevant to grounding accuracy in text-based RAG.

The number of parameters in the model.

Why wrong: Parameter count does not directly translate to grounding accuracy.

Increasing the vector dimension size to maximum

Why wrong: Dimension size is fixed by the embedding model and doesn't directly dictate relevance.

Using only fixed-size overlap chunking

Why wrong: Fixed-size chunking often splits context, reducing retrieval quality.

Amazon CloudWatch Metrics for Bedrock.

Why wrong: Provides operational metrics, not qualitative model evaluation.

AWS Glue DataBrew.

Why wrong: Used for data cleaning, not model evaluation.

Amazon SageMaker Model Monitor.

Why wrong: Used for monitoring deployed custom models, not Bedrock foundational models.

Configure the Knowledge Base to use an S3 Gateway Endpoint only.

Why wrong: While S3 needs a gateway endpoint, the communication between Bedrock and OpenSearch must also be private.

Use a NAT Gateway to route all traffic to the Bedrock service.

Why wrong: NAT Gateways are used for outbound traffic to the internet, not for secure private communication with AWS services.

Use an API Gateway with a WAF to restrict access to the vector store.

Why wrong: API Gateway acts as an entry point but does not provide the private network connectivity required for internal VPC architecture.

Use Lambda authorizers to filter the retrieved results after they are returned by the model.

Why wrong: Filtering should happen during retrieval to ensure context window efficiency and security.

Create separate vector indexes for every user role within the Knowledge Base configuration.

Why wrong: This is not a supported configuration for Amazon Bedrock Knowledge Bases.

Implement fine-grained access control using IAM policies on the S3 bucket containing the source data.

Why wrong: IAM policies control access to the bucket, not specific search results within a vector database.

Manually encrypt all files using AWS KMS before uploading to S3.

Why wrong: While good practice, it does not address PII redaction requirements.

Enable Amazon Bedrock Data Protection policies to automatically strip PII during vectorization.

Why wrong: Bedrock does not have an automatic PII-stripping feature for source documents.

Configure the Bedrock Knowledge Base to use a custom encryption key for PII filtering.

Why wrong: Encryption keys protect data at rest but do not provide logical redaction or masking of PII.

Enable Amazon Macie on the S3 bucket and configure Bedrock to ignore flagged objects.

Why wrong: Macie identifies PII but does not perform redaction as part of an automated ingestion pipeline.

Use Amazon Bedrock Guardrails to mask PII during the inference request.

Why wrong: Guardrails filter content post-generation or pre-processing but do not redact PII from the vector database ingestion pipeline.

AWS WAF on the Bedrock API endpoint

Why wrong: AWS WAF is for web traffic protection, not model input content filtering.

Amazon Macie on the S3 bucket housing the source data

Why wrong: Macie discovers PII but does not perform real-time redaction for model prompts.

AWS Glue DataBrew

Why wrong: DataBrew is for data transformation, not real-time LLM input filtering.

Semantic chunking based on paragraph breaks.

Why wrong: Semantic chunking is effective but often less precise for highly technical documents than hierarchical approaches.

Fixed-size chunking with 100 token overlap.

Why wrong: Fixed-size often leads to loss of context in technical legal documents.

Default chunking strategy.

Why wrong: The default strategy is often too simplistic for dense domain-specific content.

Knowledge Base Vector Index policies

Why wrong: These policies manage access to the index, not the conversational content.

Cross-region inference

Why wrong: This feature is for managing throughput and availability, not content policy.

Model Invocation Logs

Why wrong: These are for monitoring, not for active content filtering.

It supports multi-modal input processing natively.

Why wrong: Text embeddings are for text, not multi-modal inputs.

It is the only model that supports English.

Why wrong: Many models support English; this is not the advantage.

It prevents the model from hallucinating.

Why wrong: Embeddings represent data; they do not inherently control the generation phase hallucination.

AWS Glue DataBrew

Why wrong: Glue DataBrew is for data preparation, not model fine-tuning.

Amazon SageMaker JumpStart

Why wrong: While JumpStart helps deploy models, Bedrock Custom Models is the specific pathway for fine-tuning within the Bedrock service family.

Amazon Kendra

Why wrong: Kendra is a search service, not a platform for model training.

AWS Config

Why wrong: Config monitors resource configuration, not runtime API traffic.

Amazon Macie

Why wrong: Macie is for data discovery, not service request logging.

Amazon CloudWatch Logs

Why wrong: While it logs application events, CloudTrail is the standard for API call auditing.

Manually write code to compare generation results with source documents.

Why wrong: The service provides this natively; manual implementation is unnecessary.

Configure the Prompt Template to force the model to output document IDs.

Why wrong: While possible, this is error-prone and not the standard way provided by Bedrock.

Enable the 'Citations' feature in the Model Evaluation console.

Why wrong: Model evaluation is for testing, not for production citation delivery.

The embedding model is too small.

Why wrong: Small models are usually faster, not slower.

The source documents are too short.

Why wrong: Short documents actually improve retrieval speed.

The Bedrock service is overloaded.

Why wrong: While possible, the vector index configuration is the primary driver of retrieval speed.

Raw binary executable files.

Why wrong: Binary files do not contain semantic text suitable for embedding.

Encrypted raw stream data.

Why wrong: Stream data must be persisted into a file format first.

Proprietary database dump files.

Why wrong: These require conversion to supported formats before ingestion.

Increase the vector dimension size.

Why wrong: Dimensionality affects accuracy but not the scope of documents searched.

Use a larger model for embedding.

Why wrong: Embedding size does not help with document volume filtering.

Reduce the chunk size.

Why wrong: Smaller chunks increase count but do not help in filtering irrelevant content.

Amazon Titan Text Express

Why wrong: Titan Text Express is for text-only tasks.

Jurassic-2 Mid

Why wrong: Jurassic models are text-based.

Cohere Command R

Why wrong: Cohere Command R is primarily a text-based model for RAG.

In the Amazon Bedrock console 'Model Evaluation' tab.

Why wrong: Evaluation is for quality, not legal terms.

By asking the model during an inference session.

Why wrong: Models are not the source of truth for legal compliance terms.

In the Amazon CloudWatch monitoring dashboard.

Why wrong: CloudWatch does not contain legal terms.