SAP-C03 · Domain 1: Architecture Design and Implementation Built for the Cloud · 25% of the exam
Task 1.2: Integrate AI and ML services into architectures.
Putting AI into an architecture: generative AI with Amazon Bedrock, including across many accounts, AI agents that call tools through Amazon Bedrock AgentCore, retrieval-augmented generation with knowledge bases and vector stores, and ML inference on Amazon SageMaker AI endpoints.
Study it
Generative AI and agents in an architecture: Amazon Bedrock, AgentCore and knowledge bases
Try each one before opening the answer. Every option is explained, with the AWS documentation page that proves it.
Question 1 · choose 1
A company routes all generative AI traffic of its 40 applications through one IAM role in a central AI account, which every application assumes to call Amazon Bedrock in us-east-1. When one application runs a large nightly job, the others receive throttling errors. The company wants each application's throughput isolated from the others, while governance stays central through organization-wide policies. Which design change meets these requirements?
AHave each application call Amazon Bedrock from its own workload account, which has its own quotas
BRequest a higher tokens-per-minute quota for the central AI account so that the nightly job fits
CCreate an application inference profile for each application in the central account
DGive the nightly job a separate IAM role in the central account so that its calls are counted apart
Show the answer and why
AHave each application call Amazon Bedrock from its own workload account, which has its own quotas
Correct
Bedrock quotas are assigned per AWS account and Region. Calling from each workload account gives every application its own quota, so one heavy job no longer starves the others.
BRequest a higher tokens-per-minute quota for the central AI account so that the nightly job fits
Incorrect
A larger shared quota can still be consumed by one application, so the applications remain coupled to each other's load.
CCreate an application inference profile for each application in the central account
Incorrect
Application inference profiles track usage and cost per application, but invocations still count against the central account's quotas.
DGive the nightly job a separate IAM role in the central account so that its calls are counted apart
Incorrect
Quotas apply to the account, not to the IAM role that makes the call, so a second role in the same account shares the same limits.
Centralizing inference behind one account also centralizes its quotas. Spreading calls across workload accounts isolates throughput, and organization policies can keep governance central.
A company keeps 6 TB of engineering documents on an Amazon FSx for NetApp ONTAP volume in us-east-1, where engineers edit them through their file shares every day. It wants an Amazon Bedrock knowledge base in us-east-1 over these documents that reflects each day's edits by the next morning. The data governance team requires the documents to stay on the file system rather than be copied into other storage, and the team will not write or run ingestion code. Which design meets these requirements?
ACopy the volume to an S3 bucket each night with a scheduled AWS DataSync task, and sync the knowledge base from that bucket
BAttach an S3 access point with an internet network origin to the volume, and use it as the knowledge base's S3 data source with nightly syncs
CAttach an S3 access point restricted to the company's VPC to the volume, and use it as the knowledge base's S3 data source with nightly syncs
DAdd a custom data source and a Lambda function that reads changed files from the share and calls IngestKnowledgeBaseDocuments
Show the answer and why
ACopy the volume to an S3 bucket each night with a scheduled AWS DataSync task, and sync the knowledge base from that bucket
Incorrect
DataSync would move only changed files, but it copies them into a separate S3 bucket, which the governance team forbids.
BAttach an S3 access point with an internet network origin to the volume, and use it as the knowledge base's S3 data source with nightly syncs
Correct
An S3 access point attached to an FSx for ONTAP volume serves the file data through the S3 API while it stays on the volume. Amazon Bedrock Knowledge Bases can ingest through it, and each sync processes only the documents added, changed or deleted since the last one.
CAttach an S3 access point restricted to the company's VPC to the volume, and use it as the knowledge base's S3 data source with nightly syncs
Incorrect
The access point must have an internet network origin so that the Amazon Bedrock service can reach it. A VPC origin accepts requests only from that VPC, so ingestion cannot read the documents.
DAdd a custom data source and a Lambda function that reads changed files from the share and calls IngestKnowledgeBaseDocuments
Incorrect
Direct ingestion through a custom data source works, but the Lambda function is ingestion code that the team would have to write and run.
S3 access points for FSx for ONTAP let S3-based services such as Amazon Bedrock Knowledge Bases use file data in place. For ingestion the access point needs an internet network origin, while Block Public Access stays enforced on these access points.
A data science account hosts an Amazon SageMaker AI real-time endpoint for fraud scoring. A payments application in another account of the same organization must call the endpoint. The security team wants the data science account to control which callers may invoke the endpoint, and no long-term credentials may be used. Which design meets these requirements?
AAttach a resource-based policy to the endpoint that allows the payments application's role to call InvokeEndpoint
BShare the endpoint with the payments account through AWS RAM
CCreate a role in the data science account that allows InvokeEndpoint and trusts the payments role, and have the application assume it
DCreate an IAM user with InvokeEndpoint permission in the data science account and store its access key in the payments account's secrets
Show the answer and why
AAttach a resource-based policy to the endpoint that allows the payments application's role to call InvokeEndpoint
Incorrect
SageMaker AI does not support resource-based policies, so there is no policy to attach to the endpoint itself.
BShare the endpoint with the payments account through AWS RAM
Incorrect
AWS RAM can share several SageMaker AI resources, such as pipelines, model package groups and feature groups, but not endpoints.
CCreate a role in the data science account that allows InvokeEndpoint and trusts the payments role, and have the application assume it
Correct
The role's trust policy, owned by the data science account, decides who may assume it, and the application gets temporary credentials through AWS STS.
DCreate an IAM user with InvokeEndpoint permission in the data science account and store its access key in the payments account's secrets
Incorrect
An IAM user's access key is a long-term credential, which is not allowed.
Because SageMaker AI has no resource-based policies, cross-account access to an endpoint goes through a role in the endpoint's account that the calling account assumes.
A support assistant uses an Amazon Bedrock knowledge base in us-east-1 whose data source is an S3 bucket in us-east-1 with 400,000 documents. The disaster recovery plan requires the assistant to answer from us-west-2 within 30 minutes of a Regional outage, using knowledge that is at most a few hours old. Which design meets these requirements?
ACreate a knowledge base in us-west-2 whose data source is the existing bucket in us-east-1
BCall the model through a cross-Region inference profile so that retrieval also continues in another Region
CReplicate the bucket to us-west-2, and create and sync a knowledge base on the replica only after an outage in us-east-1 has started
DReplicate the bucket to us-west-2 with Cross-Region Replication and keep a standby knowledge base on the replica, synced on a schedule
Show the answer and why
ACreate a knowledge base in us-west-2 whose data source is the existing bucket in us-east-1
Incorrect
An S3 data source must be in the same Region as the knowledge base, so a us-west-2 knowledge base cannot use the us-east-1 bucket.
BCall the model through a cross-Region inference profile so that retrieval also continues in another Region
Incorrect
Cross-Region inference routes model invocation requests. The knowledge base, its vector store and its data source stay in us-east-1.
CReplicate the bucket to us-west-2, and create and sync a knowledge base on the replica only after an outage in us-east-1 has started
Incorrect
A first sync must parse, chunk, embed and index every one of the 400,000 documents before the knowledge base can answer, which is not a step to start inside a 30-minute recovery window.
DReplicate the bucket to us-west-2 with Cross-Region Replication and keep a standby knowledge base on the replica, synced on a schedule
Correct
The replica bucket satisfies the same-Region rule for a us-west-2 data source, and scheduled incremental syncs keep the standby knowledge base ready and only hours behind.
A knowledge base and its S3 data source are Regional. A warm standby in a second Region needs its own bucket copy and its own knowledge base that is synced ahead of time.
An agent hosted on Amazon Bedrock AgentCore Runtime has two tools: an internal inventory API that runs in private subnets of a VPC, and a public weather SaaS API on the internet. The team configured the runtime in VPC network mode with two public subnets whose route tables point 0.0.0.0/0 to an internet gateway. Calls to the inventory API work, but every call to the SaaS API times out. Which change fixes the SaaS calls and keeps the inventory access?
AConfigure the runtime with private subnets whose default route points to a NAT gateway
BAssociate Elastic IP addresses with the network interfaces that AgentCore created in the public subnets
CChange the runtime from VPC network mode to public network mode so that it reaches both APIs
DCreate an interface VPC endpoint for the AgentCore data plane in the public subnets
Show the answer and why
AConfigure the runtime with private subnets whose default route points to a NAT gateway
Correct
AgentCore gives its network interfaces private IP addresses only, so a public subnet does not give internet access. Private subnets with a route to a NAT gateway do, and the runtime stays in the VPC for the inventory API.
BAssociate Elastic IP addresses with the network interfaces that AgentCore created in the public subnets
Incorrect
The network interfaces are created and managed by AgentCore through its service-linked role and carry private addresses. The documented way to reach the internet from VPC mode is a NAT gateway, not public addresses on those interfaces.
CChange the runtime from VPC network mode to public network mode so that it reaches both APIs
Incorrect
Public mode gives internet access, but the runtime would lose the VPC connectivity it needs to reach the inventory API in private subnets.
DCreate an interface VPC endpoint for the AgentCore data plane in the public subnets
Incorrect
An AgentCore interface endpoint gives clients in the VPC private access to AgentCore APIs. It does not carry the agent's outbound calls to an internet API.
An agent in VPC mode reaches private resources through its network interfaces, which have private IP addresses. For internet tools, place it in private subnets that route to a NAT gateway, and use VPC endpoints for AWS services.
A claims application calls an Amazon SageMaker AI real-time endpoint that serves a computer vision model on one GPU instance. The model's VPC configuration lists a single private subnet. During an Availability Zone impairment last month, the endpoint stopped answering for two hours. The company wants inference to keep working if one Availability Zone fails, in the same Region, at the lowest additional cost. Which change meets these requirements?
AMove the model to a SageMaker Serverless Inference endpoint so that AWS manages the capacity across zones
BAdd an auto scaling policy to the endpoint with a minimum of one instance and a maximum of four
CDeploy a copy of the endpoint in a second Region and fail over to it with Route 53 health checks
DRun at least two instances and set the model's VPC subnets in two different Availability Zones
Show the answer and why
AMove the model to a SageMaker Serverless Inference endpoint so that AWS manages the capacity across zones
Incorrect
Serverless Inference does not support GPUs or VPC configuration, so it cannot host this model the way the application uses it.
BAdd an auto scaling policy to the endpoint with a minimum of one instance and a maximum of four
Incorrect
With a minimum of one instance in a single subnet, the endpoint still depends on one zone at the moment that zone fails.
CDeploy a copy of the endpoint in a second Region and fail over to it with Route 53 health checks
Incorrect
A second Region protects against more than a zone failure but costs a full second deployment and failover routing, which exceeds the stated need at the lowest cost.
DRun at least two instances and set the model's VPC subnets in two different Availability Zones
Correct
SageMaker AI distributes an endpoint's instances across Availability Zones when it can, and in a VPC it needs subnets in at least two zones to do so. Two instances keep the endpoint serving if one zone fails.
For production endpoints, AWS recommends multiple instances and, in a VPC, subnets in at least two Availability Zones, so that an instance failure or zone outage does not take the endpoint down.