AIP-C01 practice quiz bank
Do not use until AFTER the cold official pretest ([[study-plan]] Phase 1). If you read these questions first, the pretest stops being a clean read, and the rollback rule depends on that read.
This bank has 60 scenario questions weighted by the exam guide domain weights, plus a 15-question AWS-fundamentals appendix for rollback bucket A. Each question cites the task statement and skill it maps to in the exam guide. Every answer was checked against a live AWS docs page on 2026-11-01, and the link in the key is that page.
| Domain | Weight | Questions |
|---|---|---|
| Domain 1: Foundation Model Integration, Data Management, and Compliance | 31% | Q1-Q20 (20) |
| Domain 2: Implementation and Integration | 26% | Q21-Q35 (15) |
| Domain 3: AI Safety, Security, and Governance | 20% | Q36-Q47 (12) |
| Domain 4: Operational Efficiency and Optimization for GenAI Applications | 12% | Q48-Q54 (7) |
| Domain 5: Testing, Validation, and Troubleshooting | 11% | Q55-Q60 (6) |
| Appendix: AWS fundamentals (bucket A) | - | F1-F15 (15) |
How to use: answer a block cold, then check the key. Tag each miss A, B, or C, using the buckets from the pretest log in study-plan.md. Every question here has a single answer. The real exam also has multiple-response questions, and this bank has none.
Exam-era caution: some AWS features changed between the beta exam and today: Bedrock Agents is now "Classic," API Gateway REST APIs can stream responses, Bedrock added implicit prompt caching, and AgentCore is new. Where a change could flip the "right" answer, the key flags it with Exam-era note. Knowledge bases changed too. Since 2026-09-30, customer-managed knowledge bases can't add new Confluence, SharePoint, Salesforce, or Web Crawler connectors, and AWS points to Amazon Bedrock Managed Knowledge Base instead (data source connectors). So this bank has no connector questions, and its knowledge base questions assume a customer-managed knowledge base on S3. If an official practice question disagrees with this bank on any flagged item, trust the official set for the exam and the docs for real work.
Domain 1: Foundation Model Integration, Data Management, and Compliance (Q1-Q20)
Q1 · Task 1.1 (skill 1.1.3)
A company has eight product teams building GenAI workloads on AWS. The architecture board wants every team reviewed the same way against AWS best practices for generative AI, with the high-risk findings and the improvement plan for each workload tracked in one place. Which approach meets this with the LEAST custom effort?
- A. Publish approved reference stacks as AWS Service Catalog products for teams to launch.
- B. Run Amazon Bedrock evaluation jobs on each workload and compare the reports.
- C. Apply the Generative AI Lens to each workload in the AWS Well-Architected Tool.
- D. Deploy AWS Config conformance packs that check each GenAI account's resources.
Q2 · Task 1.2 (skill 1.2.3)
A healthcare SaaS company calls an Anthropic Claude model on Amazon Bedrock from us-east-1. During daily peaks, requests fail with ThrottlingException. A compliance rule requires that all inference stays inside the United States. Which solution meets the requirement with the LEAST operational overhead?
- A. Buy Provisioned Throughput in us-east-1 and us-west-2 and route with Route 53.
- B. Deploy the application again in us-west-2 and fail over to it on throttling.
- C. Invoke the model through a global cross-Region inference profile.
- D. Invoke the model through a US geographic cross-Region inference profile.
Q3 · Task 1.2 (skill 1.2.2)
Lambda functions behind Amazon API Gateway call Amazon Bedrock. The team wants to change the model ID without redeploying code, roll each change out gradually, and roll back automatically if a CloudWatch alarm fires. Which solution meets ALL of these requirements?
- A. Store the model ID in AWS AppConfig and deploy changes with a deployment strategy and alarm.
- B. Store the model ID in Systems Manager Parameter Store and read it on every invocation.
- C. Store the model ID in a DynamoDB table and cache it in memory with a short TTL.
- D. Store the model ID in a Lambda environment variable and update the function configuration each time.
Q4 · Task 1.2 (skill 1.2.3)
An AWS Step Functions workflow calls a primary model on Amazon Bedrock. When the primary model keeps failing, the workflow must call a secondary model instead of failing the execution. The team wants the LEAST custom code. What should they do?
- A. Add a Choice state that reads a DynamoDB health flag that a scheduled canary function updates.
- B. Move the call into a Lambda function that retries the primary model and then calls the secondary.
- C. Use a Parallel state that calls both models every time and keeps the primary result if present.
- D. Add a Retry with MaxAttempts and BackoffRate, plus a Catch that goes to a secondary-model Task.
Q5 · Task 1.2 (skill 1.2.4)
An ML team fine-tunes domain models in Amazon SageMaker AI every month. The team needs to track each model version, require an approval before a version reaches production, deploy approved versions from CI/CD, and roll back to the previous version. Which option fits BEST?
- A. Amazon ECR image tags, with a CodePipeline manual approval action.
- B. SageMaker Model Registry model groups, with versions and approval status.
- C. Amazon S3 object versioning, with an approval tag that the pipeline checks.
- D. SageMaker Model Cards, with a risk rating recorded for each model version.
Q6 · Task 1.2 (skill 1.2.4)
A team fine-tuned a Llama 3.1 8B model in SageMaker AI and saved the weights in Hugging Face safetensors format in Amazon S3. The model must be called on demand through Amazon Bedrock APIs, and the team does not want to manage any endpoints. Which approach fits BEST?
- A. Deploy the model to a SageMaker AI serverless inference endpoint.
- B. Run the model on Amazon ECS on Fargate behind a load balancer with an open-source inference server.
- C. Import the model with Bedrock Custom Model Import and call it with InvokeModel.
- D. Host the model on a SageMaker AI real-time endpoint behind API Gateway.
Q7 · Task 1.3 (skill 1.3.3)
An application must support Anthropic Claude, Meta Llama, and Amazon Nova models on Amazon Bedrock. Developers want a single request format for multi-turn chat, a system prompt, and tool use, with the least model-specific code. What should they use?
- A. Amazon Bedrock Flows, with one prompt node for each model provider.
- B. InvokeModel behind a Lambda adapter that maps one schema to each provider.
- C. InvokeModel, with each provider's native request body in the code.
- D. The Converse API, with messages, a system field, and inferenceConfig.
Q8 · Task 1.3 (skill 1.3.2)
An insurer has thousands of recorded call-center calls in Amazon S3. It wants a text-only foundation model on Amazon Bedrock to summarize each call and extract claim details. What should happen before the model call, with the LEAST custom work?
- A. Run Amazon Comprehend on each recording to pull out entities for the model.
- B. Pass each recording through an Amazon Lex bot so that the audio is converted into text.
- C. Transcribe each recording with Amazon Transcribe, then send the text to the model.
- D. Send each audio file to the model in a Converse API document content block.
Q9 · Task 1.3 (skill 1.3.1)
Each night, a pipeline lands product-catalog CSV files in Amazon S3, and those files feed a RAG corpus. Before ingestion, the description column must be complete and SKUs must be unique, and bad files must not reach the corpus. The team wants a serverless check with minimal custom code. What should they use?
- A. SageMaker Data Wrangler data quality and insights reports, run on each new file.
- B. Amazon Macie discovery jobs that flag files with problems in the bucket.
- C. AWS Glue Data Quality rules in DQDL, run in the Glue job before ingestion.
- D. A Lambda function that loads each file with pandas and runs custom checks.
Q10 · Task 1.3 (skill 1.3.4)
Millions of plain-text English support tickets a day must have people, organizations, dates, and product names extracted as metadata for retrieval filters. Each extracted value must come with a confidence score from the service, so low-confidence values can be dropped. The team cannot label data or train models. What should they use?
- A. A Bedrock model prompted to return the entities as JSON.
- B. A custom entity recognizer trained in Amazon Comprehend.
- C. Amazon Textract queries.
- D. Built-in entity detection in Amazon Comprehend.
Q11 · Task 1.4 (skill 1.4.2)
One Amazon Bedrock knowledge base indexes S3 documents from many departments. Each user may retrieve only documents from their own department and from the current year. Which solution has the LEAST operational overhead?
- A. Add a metadata.json file per document with department and year, then filter in Retrieve.
- B. Encode the department in each S3 prefix and instruct the model to use only matching sources.
- C. Create one knowledge base per department and route queries by the user's group.
- D. Retrieve without filters and drop results in Lambda based on each source URI.
Q12 · Task 1.4 (skill 1.4.5)
An Amazon Bedrock knowledge base indexes 2 million S3 documents. A few hundred of them change each day. The team must keep the knowledge base current at the lowest cost. What should they do?
- A. Re-create the data source each night so that the index matches the bucket.
- B. Turn on S3 versioning for the bucket.
- C. Run a nightly Lambda job that re-embeds every document into the vector index.
- D. Run a data source sync (StartIngestionJob) on a schedule.
Q13 · Task 1.5 (skill 1.5.1)
A knowledge base holds long technical manuals. Small chunks match queries well, but the model then lacks the surrounding section context it needs to answer. Which chunking strategy meets both needs?
- A. Hierarchical chunking.
- B. Fixed-size chunking with 50% overlap.
- C. Semantic chunking.
- D. Default chunking of about 300 tokens.
Q14 · Task 1.5 (skill 1.5.1)
An upstream process already splits a FAQ corpus into one file per question-and-answer pair. Each file must stay a single chunk in an Amazon Bedrock knowledge base, regardless of file length. Which chunking option should the team choose?
- A. Semantic chunking.
- B. Hierarchical chunking with one parent chunk per file.
- C. Fixed-size chunking at the maximum token size.
- D. No chunking for the data source.
Q15 · Task 1.5 (skill 1.5.2)
A RAG system stores Amazon Titan Text Embeddings V2 vectors at the default 1,024 dimensions in OpenSearch Serverless. The index is growing large and expensive. The team accepts a small accuracy loss but must keep the same embedding model, vector store, and chunking. What should they do?
- A. Lower numberOfResults in each Retrieve request.
- B. Add a reranker so that fewer vectors are searched.
- C. Switch the knowledge base search type to HYBRID.
- D. Re-embed the corpus at 512 or 256 dimensions.
Q16 · Task 1.5 (skill 1.5.4)
Users search an Amazon Bedrock knowledge base for exact part numbers such as XJ-4471-B. Semantic search often misses them. The vector store is Amazon OpenSearch Serverless with a filterable text field. What is the MOST effective change?
- A. Set overrideSearchType to HYBRID in the retrieval configuration.
- B. Re-embed the corpus with a higher-dimension embedding model.
- C. Switch the data source to semantic chunking to keep the IDs together.
- D. Add a reranker model to the Retrieve call to reorder the results.
Q17 · Task 1.5 (skill 1.5.4)
A RAG application retrieves 20 chunks per query, and many are only loosely related. That raises generation cost and dilutes the answers. The team wants fewer, more relevant chunks in the prompt, with the least custom code. What should they do?
- A. Raise numberOfResults and summarize the retrieved chunks with a model before generation.
- B. Add a Bedrock reranker model to the Retrieve call and keep only the top results.
- C. Turn on query decomposition in the knowledge base orchestration settings.
- D. Switch the data source to hierarchical chunking with larger parent chunks.
Q18 · Task 1.5 (skill 1.5.5)
Users ask compound questions such as "Which plan has the lower deductible, and does it cover dental?" Retrieval returns context for only one part of the question. Which change needs the LEAST custom code?
- A. Write a Lambda function that splits each question into clauses before retrieval.
- B. Turn on query decomposition in the RetrieveAndGenerate settings.
- C. Switch to hierarchical chunking with larger parent chunks for more context.
- D. Build a Step Functions Parallel state that retrieves once for each clause.
Q19 · Task 1.6 (skill 1.6.3)
Prompt engineers keep changing prompts. Production must use a fixed, approved prompt (the text plus its model and inference settings) and must be able to switch back quickly to an earlier approved prompt. Which approach needs the LEAST custom code?
- A. Keep each prompt as an AppConfig free-form configuration and deploy it on approval.
- B. Keep prompt templates in S3 with object versioning and load them by version ID.
- C. Create a version in Bedrock Prompt Management and reference it in production.
- D. Point production at the prompt's working draft in Bedrock Prompt Management.
Q20 · Task 1.6 (skill 1.6.6)
Inside the Amazon Bedrock console, and reusing prompts from Prompt Management, a team wants a visual, low-code flow: classify a ticket with a prompt; if it is a billing ticket, query a knowledge base; otherwise, call a Lambda function. Which feature fits?
- A. An AWS Step Functions workflow designed in Workflow Studio with Bedrock InvokeModel tasks.
- B. An Amazon Lex bot that routes each intent to a Lambda fulfillment function.
- C. A Bedrock agent with one action group and an associated knowledge base.
- D. Amazon Bedrock Flows, with prompt, condition, knowledge base, and Lambda nodes.
Domain 2: Implementation and Integration (Q21-Q35)
Q21 · Task 2.1 (skill 2.1.1)
A new agent built with Strands Agents must keep short-term conversation context and must also learn each user's preferences from past conversations automatically. The team does not want to build or run storage or extraction infrastructure. What should they use?
- A. A DynamoDB table of conversation turns that is loaded into each prompt.
- B. An ElastiCache for Valkey cluster plus a scheduled summarization function.
- C. Bedrock Agents session attributes, passed with each InvokeAgent call.
- D. Amazon Bedrock AgentCore Memory, with short-term and long-term memory.
Q22 · Task 2.1 (skill 2.1.7)
A company has internal REST APIs with OpenAPI specifications and several Lambda functions. Several agents need to use them as MCP tools, with authentication on inbound calls. The company wants the least custom code to build and run. What should it do?
- A. Define Bedrock Agents action groups from the OpenAPI schemas instead.
- B. Host one Lambda-based MCP server per API behind Amazon API Gateway.
- C. Run one Amazon ECS service that wraps all of the APIs as MCP tools.
- D. Create an AgentCore Gateway with OpenAPI and Lambda targets.
Q23 · Task 2.1 (skill 2.1.5)
An agentic workflow drafts customer refunds. A refund above a set limit needs a manager's approval before the workflow continues, and the approval can take up to two days. Which design fits BEST?
- A. A Standard workflow Task using .waitForTaskToken, resumed by SendTaskSuccess.
- B. An Express workflow with a Wait state loop that polls the approvals table for the decision.
- C. A Lambda function that polls an approvals table until a decision appears.
- D. An SQS queue whose consumer holds each message until an approval arrives.
Q24 · Task 2.1 (skill 2.1.3)
A ReAct-style agent loop runs in AWS Step Functions. Sometimes a tool call hangs and the execution never finishes. The team needs a time limit on each tool call and a graceful end when the limit is hit. What should they do?
- A. Add a Retry on States.ALL with MaxAttempts set to 3 on each tool Task.
- B. Raise the tool Lambda functions' timeout to the 15-minute maximum.
- C. Set TimeoutSeconds on the tool Task states and catch States.Timeout.
- D. Turn on X-Ray tracing to find hung tool calls and alert on them.
Q25 · Task 2.1 (skill 2.1.3)
A company already runs an Amazon Bedrock agent with an action group that can delete customer records. It worries that instructions hidden in retrieved content could make the agent call the delete action when the user did not intend it. Which control addresses this MOST directly?
- A. Turn on user confirmation for the delete action.
- B. Add a "confirm first" agent instruction.
- C. Attach a guardrail with the prompt attack filter to the agent.
- D. Configure the delete action group to return control to the application.
Q26 · Task 2.2 (skill 2.2.1)
A workload sends a steady, predictable, high volume of requests to one model on Amazon Bedrock around the clock, and will for at least six months. On-demand calls are throttled. The team wants reserved capacity at the lowest hourly rate. What should they buy?
- A. On-demand use through a global inference profile.
- B. Provisioned Throughput on 1-month terms, renewed monthly.
- C. Provisioned Throughput with no commitment term.
- D. Provisioned Throughput with a 6-month commitment.
Q27 · Task 2.2 (skill 2.2.3)
An assistant gets a mix of simple and complex questions. The team wants to send each request to the best-suited model in one model family to balance cost and quality, without building or training its own classifier. What should they use?
- A. An Amazon Bedrock intelligent prompt router.
- B. AppConfig model-ID flags per client.
- C. A Step Functions Choice state that routes on prompt length.
- D. A SageMaker AI classifier endpoint that picks the model.
Q28 · Task 2.2 (skill 2.2.2)
A team must self-host a 70B open-weight LLM on Amazon SageMaker AI, with tensor parallelism across GPUs and an optimized serving stack. They want the least custom container work. What should they use?
- A. A custom container on a PyTorch base image with a hand-written server.
- B. An AWS-provided LMI container on a GPU real-time endpoint.
- C. Amazon Bedrock Custom Model Import with tensor parallelism configured.
- D. A SageMaker AI serverless inference endpoint at its maximum memory size.
Q29 · Task 2.3 (skill 2.3.2)
When a contract PDF lands in Amazon S3, a summarization workflow must start. Other teams will add their own consumers of the same event later, and the bucket's configuration should not have to change each time. Which design fits BEST?
- A. Add an S3 event notification that invokes the summarization Lambda function.
- B. Turn on EventBridge for the bucket and route Object Created events with rules.
- C. Log PutObject with CloudTrail data events and trigger a Lambda function from them.
- D. Generate an hourly S3 Inventory report and process the newly listed object keys.
Q30 · Task 2.3 (skill 2.3.4)
A regulated manufacturer must keep data processing inside its own data center, but it wants to use the same AWS APIs, tools, and services as in an AWS Region. Which service fits?
- A. Amazon EC2 in the nearest Region, reached through AWS PrivateLink.
- B. AWS Wavelength.
- C. Amazon Bedrock with a geographic cross-Region inference profile.
- D. AWS Outposts.
Q31 · Task 2.3 (skill 2.3.3)
A mobile app's users sign in through a corporate SAML provider or a social provider. The app must call the AWS SDK directly with temporary, per-user credentials that are limited to approved Bedrock models. No long-term keys may ship in the app. What should the team use?
- A. STS AssumeRole called from the app with a shared IAM user's access keys.
- B. A Cognito identity pool that maps provider tokens to a scoped IAM role.
- C. A Cognito user pool that issues JWTs the app sends directly to Bedrock.
- D. IAM Identity Center permission sets assigned to each mobile app user.
Q32 · Task 2.4 (skill 2.4.3)
A batch processor uses one SDK client to call one Amazon Bedrock model at high volume, and it often gets throttled. The team wants the SDK itself to lower its request rate when throttling starts, including for first attempts. Which setting should they use?
- A. Standard retry mode plus a fixed sleep between calls.
- B. Standard retry mode with a higher max_attempts value.
- C. Adaptive retry mode.
- D. Legacy retry mode.
Q33 · Task 2.4 (skill 2.4.2)
A chatbot runs behind an existing Amazon API Gateway REST API with a Lambda proxy integration. Users wait for the full answer, and some answers take longer than 29 seconds. The team wants a faster time to first byte with the fewest changes to the existing API. What should they do?
- A. Return a job ID and have clients poll a status endpoint.
- B. Move the chatbot to a new API Gateway WebSocket API.
- C. Request an integration timeout increase above 29 seconds.
- D. Set the integration's response transfer mode to STREAM.
Q34 · Task 2.4 (skill 2.4.1)
Document-analysis requests take 2-5 minutes each, and clients can poll for results. Traffic spikes exceed the Amazon Bedrock quotas. The team wants to separate taking requests from processing them and to smooth the load. Which design fits BEST?
- A. An SNS topic that fans out each incoming request to several Lambda functions that call Bedrock.
- B. One Bedrock batch inference job submitted for each incoming request.
- C. API Gateway to SQS, a Lambda consumer that calls Bedrock, results in DynamoDB.
- D. A synchronous Lambda integration with SDK adaptive retries for throttling.
Q35 · Task 2.5 (skill 2.5.3)
Insurance claims arrive as PDFs, photos, and voice notes. The team needs the same structured fields extracted from all three formats, with minimal custom code and without hand-writing a parser for each file type. What should they use?
- A. Amazon Textract.
- B. Textract plus Transcribe, with one Lambda function for each file type.
- C. Amazon Bedrock Data Automation.
- D. A Bedrock multimodal model prompted to return JSON for each file.
Domain 3: AI Safety, Security, and Governance (Q36-Q47)
Q36 · Task 3.1 (skill 3.1.5)
An app calls InvokeModel with an Amazon Bedrock guardrail that has the prompt attack filter turned on. The developer's system prompt and the user's text are joined into one prompt string. Jailbreak attempts get through. What is the MOST likely fix?
- A. Run ApplyGuardrail on the model output as well.
- B. Raise the prompt attack filter strength to High.
- C. Add a denied topic that describes jailbreak attempts.
- D. Wrap the user's text in the guardrail input tags.
Q37 · Task 3.1 (skill 3.1.2)
A bank's chatbot must not give investment advice, whether a user asks for it or the model starts to offer it. Normal account questions must still work. Which Amazon Bedrock Guardrails feature fits BEST?
- A. A word filter for terms such as "buy", "stock", and "portfolio".
- B. A content filter for the Misconduct category, set to High strength for inputs and outputs.
- C. A denied topic for investment advice, with a definition and examples.
- D. A contextual grounding check against the bank's approved FAQ pages.
Q38 · Task 3.1 (skill 3.1.3)
A RAG application sometimes returns facts that are not in the retrieved passages, or answers that don't address the question. For each request, the team wants to block responses that aren't supported by that request's retrieved passages. What should they configure?
- A. A reranker model that removes low-relevance chunks before the generation step.
- B. Automated Reasoning checks built from the company's policy documents.
- C. A contextual grounding check with grounding and relevance thresholds.
- D. A second model call that asks the model to verify its own answer.
Q39 · Task 3.2 (skill 3.2.3)
A call-summary feature must keep its summaries readable while replacing customer names and email addresses. Summaries that contain PII must still be produced, not rejected. Which Amazon Bedrock Guardrails setting meets this?
- A. A custom regex filter set to Block for email address patterns.
- B. A sensitive information filter set to Mask for NAME and EMAIL.
- C. Amazon Comprehend PII detection that rejects transcripts with PII.
- D. A sensitive information filter set to Block for NAME and EMAIL.
Q40 · Task 3.2 (skill 3.2.2)
A team uses an Amazon Bedrock guardrail with PII masking and sends model invocation logs to CloudWatch Logs. An auditor finds unmasked PII in the input fields of the logs. What is the correct explanation and fix?
- A. Send the invocation logs to Amazon S3 instead of CloudWatch Logs for storage.
- B. This is expected; add a CloudWatch Logs data protection policy.
- C. Apply the guardrail to model inputs as well as to outputs.
- D. Switch the guardrail's PII action from Mask to Block.
Q41 · Task 3.1 (skill 3.1.4)
An application uses a model hosted outside Amazon Bedrock. The team wants to reuse the denied topics, content filters, and PII rules already defined in a Bedrock guardrail, and wants to check user input before retrieval. What should they do?
- A. Rebuild the rules as Amazon Comprehend custom classifiers.
- B. Route the outside model's traffic through a Bedrock inference profile.
- C. Call the ApplyGuardrail API on user input and on model output.
- D. Attach the guardrail to the RetrieveAndGenerate call only.
Q42 · Task 3.2 (skill 3.2.1)
Lambda functions in private subnets with no internet path must call Amazon Bedrock Runtime. Traffic must not cross the public internet, and the team wants to limit what can be done through that connection. What should they set up?
- A. An interface endpoint for the bedrock control-plane service with a policy.
- B. A bedrock-runtime interface endpoint with private DNS and an endpoint policy.
- C. A gateway endpoint for Amazon Bedrock added to the private route tables.
- D. A NAT gateway with an egress rule limited to Bedrock's published IP ranges.
Q43 · Task 3.2 (skill 3.2.1)
A security team must stop every account in the organization from running inference on any foundation model except an approved list. The control must cover every inference API, including Converse, and must apply to roles created later. Which control fits BEST?
- A. An SCP that denies only bedrock:Converse and bedrock:ConverseStream.
- B. A VPC endpoint policy on every account's Bedrock runtime interface endpoint.
- C. An SCP that denies bedrock:InvokeModel* on unapproved model ARNs.
- D. An IAM permissions boundary attached to each developer role in each account.
Q44 · Task 3.2 (skill 3.2.2)
Before a large S3 corpus is added to a knowledge base, compliance wants an automated report of which objects contain PII, with as little custom code as possible. Which service fits?
- A. Amazon GuardDuty S3 Protection turned on for the corpus bucket.
- B. Amazon Macie sensitive data discovery on the bucket.
- C. A guardrail sensitive information filter applied during ingestion.
- D. A Lambda function that sends each object to Comprehend PII detection.
Q45 · Task 3.3 (skill 3.3.2)
An auditor asks for two things: a record of who called which Amazon Bedrock model and when, and the full prompt and response text of each call. Which combination provides both?
- A. CloudTrail management events, plus model invocation logging.
- B. AWS X-Ray traces, plus CloudWatch AWS/Bedrock metrics.
- C. CloudTrail data events, plus CloudWatch Logs Insights.
- D. CloudTrail management events, plus CloudTrail data events.
Q46 · Task 3.3 (skill 3.3.1)
A regulator wants documentation of a fine-tuned model's intended use, risk rating, training details, and evaluation results, and wants it kept up to date across model versions. What should the team use?
- A. Amazon SageMaker Model Cards.
- B. SageMaker Model Registry custom metadata key-value pairs.
- C. Amazon Bedrock evaluation job reports stored in S3.
- D. AWS Glue Data Catalog.
Q47 · Task 3.4 (skill 3.4.1)
Users ask why an Amazon Bedrock agent took an action. Developers need the agent's step-by-step reasoning plus the inputs and outputs of each action group and knowledge base query, both to explain behavior and to troubleshoot. What should they use?
- A. Query CloudTrail data events recorded for InvokeAgent calls.
- B. Turn on trace in InvokeAgent and read the orchestration trace.
- C. Turn on model invocation logging and read the agent's prompts.
- D. Turn on X-Ray active tracing for the action group Lambda functions.
Domain 4: Operational Efficiency and Optimization for GenAI Applications (Q48-Q54)
Q48 · Task 4.1 (skill 4.1.4)
In a document Q&A app, users ask many questions about the same 40-page contract within a few minutes, and every answer must consider the whole contract. Input tokens for the contract dominate the cost. Which change MOST reduces cost and latency?
- A. Prompt caching, with the contract as a static prefix.
- B. Provisioned Throughput sized to the peak token rate.
- C. A knowledge base over the contract with a reranker model.
- D. A DynamoDB cache keyed on a hash of each exact question.
Q49 · Task 4.1 (skill 4.1.3)
Each night, a team must rewrite 500,000 product descriptions with a model on Amazon Bedrock. Results are due by morning, and there is no real-time need. Which option is MOST cost-effective?
- A. On-demand calls through a global cross-Region inference profile.
- B. Parallel on-demand Converse calls from a Step Functions Distributed Map.
- C. A Bedrock batch inference job over input files in Amazon S3.
- D. Provisioned Throughput with no commitment, for the nightly window only.
Q50 · Task 4.2 (skill 4.2.1)
A voice assistant uses a model that supports latency-optimized inference on Amazon Bedrock. The team needs lower total response time, not just a faster first token. They can't fine-tune, and they want minimal code changes. What should they do?
- A. Set the latency setting to optimized in the runtime call.
- B. Buy Provisioned Throughput for the model.
- C. Call the model through a global inference profile.
- D. Switch to ConverseStream with standard inference.
Q51 · Task 4.3 (skill 4.3.2)
Operations wants alarms on throttled calls and on token usage for each Amazon Bedrock model, with the least setup. What should they use?
- A. Metric filters on the model invocation log group for throttles.
- B. CloudWatch alarms on AWS/Bedrock throttle and token metrics.
- C. A scheduled Logs Insights query over CloudTrail events.
- D. X-Ray groups with alarms on throttled trace segments.
Q52 · Task 4.3 (skill 4.3.2)
Token usage follows a daily pattern. The team wants alerts when usage moves away from that normal pattern, without maintaining static thresholds. What should they configure?
- A. A static CloudWatch alarm at twice the trailing average.
- B. A CloudWatch anomaly detection alarm on the token metric.
- C. AWS Cost Anomaly Detection on the Amazon Bedrock service.
- D. A metric math alarm that compares to last week's values.
Q53 · Task 4.3 (skill 4.3.2)
Finance wants alerts when Amazon Bedrock spend jumps unexpectedly. Spend grows every month as adoption grows, so any fixed threshold either fires all the time or misses spikes. Alerts must reach Slack through an SNS topic, and finance wants the likely root cause ranked by cost impact. What should the team set up?
- A. A CloudWatch billing alarm on EstimatedCharges with a monthly threshold.
- B. A CloudWatch anomaly detection alarm on InputTokenCount that notifies SNS.
- C. An AWS Cost Anomaly Detection monitor with an SNS alert subscription.
- D. A scheduled Lambda function that compares Cost Explorer data and posts to SNS.
Q54 · Task 4.2 (skill 4.2.4)
A field-extraction prompt gives different outputs for identical inputs. The team must keep the same model and prompt, must not add latency, and must not cut off long outputs. Which change helps MOST?
- A. Lower maxTokens so the model has less room to vary.
- B. Call the model three times and keep the majority answer.
- C. Raise Top K so the model considers more candidate tokens.
- D. Lower the temperature, and lower Top P if needed.
Domain 5: Testing, Validation, and Troubleshooting (Q55-Q60)
Q55 · Task 5.1 (skill 5.1.2)
A team wants to compare a model hosted outside AWS with Amazon Bedrock models on a custom quality metric across thousands of prompts. It already has the outside model's responses and cannot call that model again. Scoring must be automated. What should it use?
- A. An LLM-as-a-judge job with its own responses in the dataset.
- B. A programmatic evaluation job with a built-in dataset.
- C. Custom Model Import of the outside model, then an evaluation.
- D. A human evaluation job that uses a private work team.
Q56 · Task 5.1 (skill 5.1.6)
A team wants to measure how well an Amazon Bedrock knowledge base retrieves relevant content, separately from how well the model writes the answers. Which evaluation fits?
- A. A retrieve-only RAG evaluation job.
- B. An LLM-as-a-judge model evaluation job.
- C. A retrieve-and-generate RAG evaluation job.
- D. A programmatic model evaluation job with a custom dataset.
Q57 · Task 5.2 (skill 5.2.1)
Long summaries from the Converse API end mid-sentence. On every cut-off response, usage.outputTokens equals the configured limit of 512, while usage.inputTokens is far below the model's context window and no guardrail trace appears. What is the cause and fix?
- A. A stop sequence matched; remove the stop sequences from the request.
- B. The output hit maxTokens; raise it or ask for shorter summaries.
- C. A guardrail cut the response; review the guardrail's content filters.
- D. The input overflowed the context window; chunk the source text first.
Q58 · Task 5.2 (skill 5.2.2)
During peak hours, logs show HTTP 429 ThrottlingException responses from Amazon Bedrock. No other error codes appear at any time. Which response is correct?
- A. Retry with backoff and jitter, check quotas, and request an increase.
- B. Fix the calling role's IAM permissions for bedrock:InvokeModel.
- C. Correct the model ID and Region used in the requests during peaks.
- D. Validate each request body against the model's schema before sending.
Q59 · Task 5.2 (skill 5.2.4)
Users report that a chatbot cites an outdated policy, even though the policy file in the knowledge base's S3 data source was updated yesterday. What is the MOST likely cause?
- A. Prompt caching is still serving the old policy text.
- B. The data source has not been synced since the update.
- C. The embedding model has drifted and needs to be replaced.
- D. numberOfResults is too low to reach the newer version.
Q60 · Task 5.2 (skill 5.2.5)
Model invocation logs go to CloudWatch Logs. For a one-time investigation, the team needs to find which prompt-template versions and models produce malformed JSON across thousands of calls, grouped by version. What is the MOST efficient tool?
- A. CloudWatch Logs Insights queries on the invocation logs.
- B. CloudTrail Lake queries on the InvokeModel events.
- C. An X-Ray service map of the calling application.
- D. A metric filter that counts JSON parse errors.
Appendix: AWS fundamentals check (rollback bucket A, F1-F15)
These aren't GenAI questions. A miss here counts as bucket A. Use them to sanity-check the A tags from the pretest.
F1 · IAM policy evaluation
A role's identity-based policy allows s3:GetObject on a bucket in the same account. The bucket policy has an explicit Deny for that role. What happens when the role calls GetObject?
- A. Allowed; a same-account identity policy is enough.
- B. Denied; an explicit deny overrides the allow.
- C. Denied only if an SCP denies it.
- D. Allowed; bucket policies apply only to other accounts.
F2 · IAM roles for Lambda
What is the recommended way to give a Lambda function permission to call Amazon Bedrock?
- A. Add a resource-based policy on the function for Bedrock.
- B. Attach the permissions to the function's execution role.
- C. Store an IAM user's access keys in Secrets Manager.
- D. Put IAM user keys in encrypted environment variables.
F3 · Service control policies
An SCP attached to an account allows bedrock:*. A role in that account has no identity-based policies. Can the role invoke a model?
- A. No; SCPs never grant permissions.
- B. Yes; the SCP allows it.
- C. Yes, but only from the management account.
- D. Only if a permissions boundary allows it too.
F4 · Gateway endpoints
EC2 instances in a private subnet read large files from Amazon S3 in the same Region. The team wants private connectivity at the lowest cost. What should it use?
- A. An S3 access point with a VPC network origin.
- B. An interface endpoint for S3 in each subnet.
- C. A NAT gateway in each Availability Zone.
- D. A gateway endpoint for S3 in the route tables.
F5 · Interface endpoint DNS
After the team adds an interface VPC endpoint for a service, existing SDK code still calls the service's public Regional endpoint name. How can that code use the endpoint with no code change?
- A. Add a Route 53 Resolver outbound endpoint to the VPC.
- B. Set the SDK endpoint URL to the endpoint's DNS name.
- C. Turn on private DNS for the endpoint.
- D. Use a Route 53 public hosted zone.
F6 · Endpoint policies
A VPC endpoint has an endpoint policy that allows full access. A role's identity-based policy denies the action. Is the request allowed through the endpoint?
- A. Yes; the endpoint policy allows full access.
- B. No; endpoint policies don't override identity policies.
- C. Yes; endpoint traffic skips identity-based policies.
- D. Yes, as long as private DNS is turned on.
F7 · KMS key policies
A team creates a customer managed AWS KMS key and plans to control access only with IAM policies. What must be true?
- A. The key has one key policy, and it must allow IAM-based access.
- B. Every principal that uses the key needs its own grant.
- C. The key needs a separate resource policy as well as a key policy.
- D. IAM policies alone control the key; the key policy is optional.
F8 · KMS key rotation
Automatic rotation is turned on for a customer managed KMS key. Which statement is true?
- A. Old ciphertext must be re-encrypted before old material is deleted.
- B. Rotation requires creating a new key and moving the alias to it.
- C. The key ID changes on each rotation, so aliases must be updated.
- D. New material yearly by default; old ciphertext still decrypts.
F9 · S3 Lifecycle
Raw prompt logs in S3 must move to a cheaper storage class after 30 days and be deleted after one year under a retention policy. What should the team use?
- A. A scheduled Lambda function that copies objects and deletes them.
- B. S3 Object Lock in compliance mode with a one-year retention period.
- C. A Lifecycle configuration with transition and expiration actions.
- D. S3 Intelligent-Tiering with the archive access tier turned on.
F10 · S3 default encryption
A team uploads objects to a new S3 bucket without setting any encryption options. What happens?
- A. S3 encrypts them with SSE-S3 by default.
- B. S3 encrypts them with the AWS managed KMS key for S3.
- C. The upload fails.
- D. They are stored unencrypted until default encryption is set.
F11 · S3 Block Public Access
At which levels can S3 Block Public Access be managed?
- A. Only at the bucket level, one bucket at a time.
- B. Account and bucket levels, but not the organization.
- C. Organization, account, and bucket or access point.
- D. At the object level, through each object's ACL.
F12 · Lambda timeout
A standard Lambda function (not Lambda Managed Instances) must run a 40-minute document-processing job. What is the problem?
- A. Functions in a VPC are limited to a 5-minute timeout.
- B. The 900-second maximum timeout is too short.
- C. The 10,240 MB memory limit stops long-running jobs.
- D. Lambda can't read S3 objects larger than 512 MB.
F13 · Lambda payload
A client sends a 20 MB document in the body of a synchronous Lambda invocation. What happens, and what is the better pattern?
- A. It fails only for functions attached to a VPC.
- B. It works if the function uses response streaming.
- C. It works if the function has 10,240 MB of memory.
- D. It fails at the 6 MB limit; pass an S3 key instead.
F14 · CloudWatch metric filters
The team wants an alarm whenever ValidationException appears in an application log group more than 10 times in 5 minutes, with the least custom code. What should it create?
- A. A CloudTrail trail with an EventBridge rule on the error.
- B. An X-Ray sampling rule that keeps every error trace.
- C. A metric filter on the log group, plus an alarm.
- D. A subscription filter to a Lambda function that counts errors.
F15 · CloudWatch Logs retention
Log storage costs keep growing. Nobody has changed any settings on the team's CloudWatch Logs log groups. Why, and what should the team do?
- A. Logs are kept forever by default; set retention.
- B. Logs expire after 30 days; the cost is Insights queries.
- C. Logs default to a per-query log class; switch to Standard.
- D. Logs move to S3 Glacier.
Answer key
Letter counts, main set: A 15 · B 15 · C 15 · D 15. Longest run of one letter: 3. Appendix: A 4 · B 4 · C 4 · D 3, longest run 2. The letters were assigned by a random shuffle.
Domain 1
- Q1 - C. In the WA Tool, lenses let you "consistently measure your architectures against best practices," and each lens comes with its own questions and improvement plan. The exam guide names the WA Tool Generative AI Lens for this skill. Service Catalog standardizes what teams provision, Config checks resource settings, and evaluations score models; none of these is a best-practice review with tracked risks. Using lenses in AWS WA Tool · Generative AI Lens
- Q2 - D. Geographic profiles keep processing inside the geography; global profiles route to any commercial Region, which breaks the residency rule. Cross-Region inference adds no routing cost and needs no new infrastructure. Inference profiles also don't support Provisioned Throughput. Cross-Region inference
- Q3 - A. AppConfig changes behavior without a code deploy, rolls changes out gradually with deployment strategies, and rolls back automatically when a CloudWatch alarm fires. Parameter Store and DynamoDB avoid a redeploy but offer neither a gradual rollout nor an alarm-based rollback. What is AWS AppConfig?
- Q4 - D. Step Functions applies retriers first. If retries don't fix the error, it moves to the catcher's
Nextstate. That gives a fallback branch with no custom code. The Lambda and canary designs work, but they need custom code, and the Parallel design pays for two calls on every request. Handling errors in Step Functions - Q5 - B. Model Registry catalogs model versions in model groups, manages approval status, deploys models to production, and automates deployment with CI/CD. Model Cards document a model; they don't gate deployment. Model Registry
- Q6 - C. Custom Model Import brings in customized open-weight models from S3 (Llama is a supported architecture) and serves them on demand through
InvokeModel. Every other option means running an endpoint or service yourself, outside the Bedrock APIs. Limits: Custom Model Import doesn't support batch inference and is offered in only a few Regions. Custom Model Import - Q7 - D. The Converse API "provides a consistent API that works with all Amazon Bedrock models that support messages," so you write the code once. It also supports tool use and guardrails. Converse API
- Q8 - C. The model only accepts text, so the audio has to become text first. Amazon Transcribe is the AWS speech-to-text service. Comprehend works only on text, and Lex is for building conversational bots, not batch transcription. What is Amazon Transcribe?
- Q9 - C. Glue Data Quality is managed and serverless, and you write its rules in DQDL. The pandas Lambda works but is all custom code. Data Wrangler is an interactive data preparation tool, not a pipeline gate. Macie finds sensitive data; it doesn't check data quality. AWS Glue Data Quality
- Q10 - D. Comprehend's built-in entity types cover people, organizations, commercial items, and dates, and "each entity also has a score that indicates the level of confidence." A custom recognizer needs training data. A prompted FM doesn't return a service-calibrated confidence score. Textract is for document text detection and analysis (Textract). Comprehend entities
- Q11 - A. To filter an S3 data source on metadata, each document needs a
fileName.extension.metadata.jsonfile; you then apply the filters at query time. A single knowledge base with filters is less work than one knowledge base per department. A prompt instruction is not an access control. Configure and customize queries Exam-era note: AWS now recommends Amazon Bedrock Managed Knowledge Base for the Confluence, SharePoint, Salesforce, and Web Crawler connectors. Managed knowledge bases don't support thestartsWithorstringContainsfilters (useequals,in, and the others). This question uses a customer-managed knowledge base on S3. - Q12 - D. "Syncing is incremental, so Amazon Bedrock only processes added, modified, or deleted documents since the last sync." Re-creating the data source or re-embedding everything pays for millions of unchanged documents. S3 versioning does not update the index. Sync your data Exam-era note: This assumes a customer-managed knowledge base. AWS docs now say: "For optimized retrieval accuracy and a managed experience, we recommend Amazon Bedrock Managed Knowledge Base," and managed knowledge bases don't support the
startsWithorstringContainsmetadata filters. - Q13 - A. Hierarchical chunking matches on small child chunks, then replaces them with their larger parent chunks, so the model gets more context. Semantic chunking and overlap change where chunks break; they don't add parent context at retrieval time. Content chunking
- Q14 - D. With no chunking, each document is one chunk; the docs suggest splitting files first and then choosing no chunking. Every other strategy has a maximum token size, so a long file could still be split. Trade-off: you lose page-number citations and page-number filtering. Content chunking
- Q15 - D. Titan Text Embeddings V2 offers output sizes of 1,024 (default), 512, and 256, and fewer dimensions mean a smaller index. The other options change what a query returns, not how much data is stored. Titan embedding models
- Q16 - A. Hybrid search adds raw-text search to vector search, which catches exact identifiers. It is supported for Amazon RDS, OpenSearch Serverless, and MongoDB vector stores that have a filterable text field. A reranker only reorders what semantic search already found. Configure and customize queries
- Q17 - B. A reranker scores each chunk's relevance to the query and reorders the results, so you can pass fewer, more relevant chunks. That also lowers cost and latency. It works inside
RetrieveandRetrieveAndGenerate. Reranker models - Q18 - B. Query decomposition is a built-in orchestration option that splits a complex query into sub-queries. The Lambda and Step Functions designs are valid patterns from the guide, but they need custom code. Configure and customize queries
- Q19 - C. A version is a point-in-time snapshot of a prompt and its configuration; you deploy versions to production and switch between them. The draft keeps changing. The S3 and AppConfig designs need code to store and apply the model and inference settings. Deploy a prompt using versions
- Q20 - D. Flows is the visual builder in Bedrock that links prompt, knowledge base, and Lambda nodes, and a condition node sends data to different nodes based on conditions. (The exam guide calls this "Prompt Flows.") Step Functions is also visual, but it isn't in the Bedrock console and has no prompt node. Node types in flows
Domain 2
- Q21 - D. AgentCore Memory is a fully managed service for both short-term context and long-term memory, and it works with agent frameworks. A DynamoDB table stores conversation turns but doesn't extract preferences. AgentCore Memory Exam-era note: Bedrock Agents is now "Amazon Bedrock Agents Classic" and is no longer open to new customers, so AgentCore is the current answer for a new build.
- Q22 - D. AgentCore Gateway turns APIs, Lambda functions, and existing services into MCP-compatible tools; it accepts OpenAPI, Smithy, and Lambda inputs and handles both inbound and outbound authentication. Lambda- or ECS-hosted MCP servers are valid patterns from the guide, but they need more code. AgentCore Gateway
- Q23 - A. With the callback pattern, Step Functions passes a task token and waits until
SendTaskSuccessorSendTaskFailurereturns it. A waiting task waits until the execution reaches the one-year quota. SetHeartbeatSeconds(the docs' recommended guard) and/orTimeoutSeconds, and catchStates.Timeout. Service integration patterns - Q24 - C. A Task state that runs past
TimeoutSecondsfails withStates.Timeout, and aCatchcan route it to a graceful end.States.Timeoutalso fires when the whole execution runs past its ownTimeoutSeconds. Task state - Q25 - A. AWS recommends user confirmation as a defense against malicious prompt injection: before calling the function, the agent asks the user to confirm or deny the action. The prompt attack filter checks tagged user input, not tool results. Return of control hands the call to the application but doesn't ask the user by itself. User confirmation Exam-era note: This is a feature of Bedrock Agents, now Bedrock Agents Classic, which existing customers still use.
- Q26 - D. Provisioned Throughput is billed by the hour, and "the longer the commitment duration, the more discounted the hourly price becomes" (no commitment, 1 month, or 6 months). A global profile is cheaper on-demand capacity, but it reserves nothing. Provisioned Throughput Exam-era note: The throttling docs now also mention a "Reserved Tier" for some models. Cross-Region inference profiles don't support Provisioned Throughput.
- Q27 - A. Intelligent prompt routing offers one serverless endpoint that predicts response quality for each request and routes it between models in the same family, balancing quality and cost. Routing on prompt length is a crude stand-in, and a classifier would have to be built and trained. Intelligent prompt routing
- Q28 - B. Large model inference (LMI) containers are AWS-provided Docker containers built for deploying and optimizing LLM inference on SageMaker AI. Serverless Inference doesn't support GPUs (serverless exclusions), and Custom Model Import isn't self-hosting on SageMaker AI. LMI container docs
- Q29 - B. Once EventBridge is on for a bucket, S3 sends every listed event type, including Object Created, and you "don't need to select which event types you want to deliver." Each new consumer is just a new rule. A direct event notification is tied to the bucket's configuration. Using EventBridge with S3
- Q30 - D. Outposts extends AWS infrastructure, services, APIs, and tools to customer premises. Wavelength is edge compute for low latency to mobile devices (Wavelength). The other two options still process data in an AWS Region. What is AWS Outposts?
- Q31 - B. Identity pools exchange federated identities for temporary AWS credentials, and IAM roles and policies set what users can do. Bedrock calls need AWS credentials, not a user pool JWT. Identity Center is for workforce access. Shared IAM user keys are long-term keys. Cognito identity pools
- Q32 - C. Adaptive mode adds a client-side rate limiter that "can delay or block the initial request, not just retries," when throttling is detected. AWS names "AI workloads that call a single API operation at high volume" as a fit, but does not recommend adaptive mode as a general default. Retry behavior
- Q33 - D. When the response transfer mode is
STREAM, API Gateway starts sending the response before it is complete. That lowers time to first byte for chatbots and gets past the 29-second timeout without a quota increase. A WebSocket API also streams, but it means building a new API. Stream the integration response Exam-era note: Older material answers this with a WebSocket API; REST response streaming is newer and needs fewer changes. - Q34 - C. With SQS, Lambda processes messages in batches; messages in a batch stay hidden for the visibility timeout. Use partial batch responses so successful messages are not retried. The queue buffers spikes, so Bedrock doesn't see them directly. SNS fan-out doesn't buffer the load. Using Lambda with SQS
- Q35 - C. Bedrock Data Automation uses generative AI to turn multimodal content (documents, images, video, and audio) into structured formats. The Textract-plus-Transcribe design needs code for each file type, and prompting a model yourself means writing and maintaining the extraction logic. Bedrock Data Automation
Domain 3
- Q36 - D. For
InvokeModelandInvokeModelWithResponseStream, "if there are no tags, prompt attacks for those use cases will not be filtered." The prompt attack filter also doesn't check tool results. Prompt attacks - Q37 - C. Denied topics block the topics you define, in both user queries and model responses. A word filter would block normal questions and still miss reworded advice. Misconduct is a fixed harm category, not a business topic. Denied topics
- Q38 - C. A contextual grounding check scores grounding (is the answer supported by the source?) and relevance (does it answer the query?), and blocks responses that score below your thresholds. Automated Reasoning checks test answers against fixed logical rules, not each request's retrieved passages. Contextual grounding check
- Q39 - B. The Mask action replaces detected PII with its type, for example
{NAME}or{EMAIL}. Block rejects the whole content, which breaks the rule that summaries must still be produced. Sensitive information filters - Q40 - B. PII masking doesn't reach invocation logs: "the input field in Amazon CloudWatch Logs always contains the original, unmodified request." A CloudWatch Logs data protection policy masks it (masking). Sensitive information filters
- Q41 - C.
ApplyGuardrailchecks any text against an existing guardrail without calling a foundation model, so you can check input before retrieval and output after generation. ApplyGuardrail API - Q42 - B. PrivateLink interface endpoints connect without an internet gateway, NAT, VPN, or Direct Connect, and they accept endpoint policies. Inference uses the
bedrock-runtimeservice, notbedrock. Gateway endpoints exist only for S3 and DynamoDB. Bedrock interface endpoints - Q43 - C. AWS's example denies
bedrock:InvokeModelonfoundation-modelARNs and says it "can be used as a service control policy (SCP)." Converse is blocked automatically when InvokeModel is denied, andbedrock:InvokeModel*covers all inference actions. A permissions boundary only covers the roles it is attached to. Identity-based policy examples - Q44 - B. Macie uses machine learning and pattern matching to find sensitive data in S3 and creates findings for it. The Comprehend design works but is custom code. Guardrails check model input and output, not stored objects. What is Macie?
- Q45 - A. CloudTrail records
InvokeModelandConverseas management events, which are logged by default. Invocation logging captures the full request and response, but it is off by default (invocation logging). CloudTrail events never contain the prompt text. CloudTrail for Bedrock - Q46 - A. Model Cards record intended use, risk rating, training details, and evaluation results in one place for governance and reporting. They can be linked to Model Registry versions, but registry metadata on its own isn't a governance document. SageMaker Model Cards
- Q47 - B. The trace shows the agent's reasoning and the inputs and outputs of action groups and knowledge bases; the orchestration trace includes a
rationale. CloudTrail records that a call happened, not the reasoning behind it. Track the agent's reasoning with trace Exam-era note: This is Bedrock Agents, now Classic. For AgentCore agents, use AgentCore Observability, which shows each step of the agent workflow and lets you inspect the execution path (AgentCore observability).
Domain 4
- Q48 - A. Prompt caching lowers latency and input-token cost for long contexts that are reused. The cache TTL resets on each hit, and many models support 5 minutes. An exact-question cache rarely hits, and RAG breaks the whole-contract requirement. Prompt caching Exam-era note: Bedrock now also offers implicit (automatic, best-effort) caching alongside explicit cache checkpoints.
- Q49 - C. Batch inference runs many prompts asynchronously, from input files in S3 to output in S3. AWS's pricing page lists batch at a 50% lower price than on-demand for select models. Global profiles save about 10%. Batch inference
- Q50 - A. Setting the "Latency" parameter to "optimized" needs no setup or fine-tuning. Streaming speeds up the first token, not the total time, and Provisioned Throughput buys capacity, not speed. Latency-optimized inference Exam-era note: The feature is still in preview and covers only a few models (for example, Nova Pro, Claude 3.5 Haiku, and Llama 3.1 405B and 70B).
- Q51 - B. The runtime endpoint publishes
InvocationThrottles,InputTokenCount,OutputTokenCount,TimeToFirstToken, and more underAWS/Bedrock, both with and without theModelIddimension. No logging setup is needed. Runtime CloudWatch metrics - Q52 - B. CloudWatch anomaly detection learns a metric's normal baseline and produces a band of expected values, so you don't need static thresholds. Cost Anomaly Detection watches spend, not token metrics. CloudWatch anomaly detection
- Q53 - C. Cost Anomaly Detection uses machine learning to detect unusual spend, sends alerts by email or SNS, and lets you investigate root causes "ranked by their dollar impact." Token-count alarms track usage, not spend, and a billing alarm uses a fixed threshold. Cost Anomaly Detection
- Q54 - D. A lower temperature "steepens the function and leads to more deterministic responses," and Top P limits the candidate tokens. A lower maxTokens truncates output, majority voting adds latency, and a higher Top K adds variety. Inference parameters
Domain 5
- Q55 - A. LLM-as-a-judge jobs accept your own inference responses (Bedrock then skips the model call) and support custom metrics. A human job isn't automated, and importing the model means calling it again. Judge-model evaluations
- Q56 - A. A retrieve-only job reports on the retrieved data alone. A retrieve-and-generate job also scores the generated response, so retrieval and generation quality get mixed together. RAG evaluations
- Q57 - B.
maxTokensis "the maximum number of tokens to allow in the generated response." Output tokens that exactly match the limit, with plenty of input headroom and no guardrail trace, point to the output limit. InferenceConfiguration - Q58 - A. A 429 means the request exceeded the account's Bedrock quotas. Use retries with exponential backoff and jitter, check the quotas, and for sustained load request an increase or use Provisioned Throughput or Reserved Tier. The other options fix 403, 400, and 404 errors. API error codes
- Q59 - B. You must sync the data source after files are added, changed, or deleted so that the knowledge base re-indexes them. Sync your data Exam-era note: This assumes a customer-managed knowledge base. AWS docs now say: "For optimized retrieval accuracy and a managed experience, we recommend Amazon Bedrock Managed Knowledge Base," and managed knowledge bases don't support the
startsWithorstringContainsmetadata filters. - Q60 - A. Logs Insights runs interactive queries over CloudWatch Logs, and the exam guide names it for analyzing prompts and responses (skill 2.5.6). CloudTrail has no prompt content, and a metric filter gives a count, not a breakdown by version. CloudWatch Logs Insights
Appendix
- F1 - B. An explicit deny in either the identity-based policy or the resource-based policy overrides the allow. Policy evaluation logic
- F2 - B. A Lambda function's execution role is an IAM role that grants the function permission to access AWS services. Lambda execution role
- F3 - A. "An SCP never grants permissions." An SCP only sets the maximum permissions, so the role still needs a policy that allows the action. SCPs
- F4 - D. Gateway endpoints serve S3 and DynamoDB through route tables, and "there is no additional charge for using gateway endpoints." An access point limits access to a VPC but doesn't provide the network path. Gateway endpoints
- F5 - C. With private DNS turned on (and DNS hostnames and DNS resolution turned on in the VPC), requests to the public endpoint name resolve to the endpoint's private IP addresses. Setting the SDK endpoint URL is a code change. Access AWS services through PrivateLink
- F6 - B. "An endpoint policy does not override or replace identity-based policies or resource-based policies." Endpoint policies
- F7 - A. "Every KMS key must have exactly one key policy." IAM policies and grants work alongside it, and only when the key policy allows them. Key policies
- F8 - D. By default, KMS generates new key material every year. You can set a custom rotation period or rotate on demand, and KMS automatically picks the right key material to decrypt older ciphertext. Rotate keys
- F9 - C. Transition actions move objects to another storage class, and expiration actions delete them. Intelligent-Tiering moves objects between tiers but never deletes them. Object lifecycle management
- F10 - A. Since January 5, 2023, S3 encrypts all new object uploads with SSE-S3 at no extra cost. Default bucket encryption
- F11 - C. You can manage Block Public Access at the organization level (AWS Organizations), the account level, and the bucket and access point level. Block Public Access
- F12 - B. The maximum function timeout is 900 seconds (15 minutes). Current-docs note: Lambda Managed Instances allow up to 90 minutes for asynchronous and event-source-mapping invocations, which is why the question says "standard." Lambda quotas
- F13 - D. Synchronous invocations are limited to 6 MB each for the request and the response. Only streamed responses can be larger, so streaming doesn't help a large request. Lambda quotas
- F14 - C. Metric filters turn log data into CloudWatch metrics that you can graph or set an alarm on. A subscription filter to Lambda works but needs custom code. Metric filters
- F15 - A. "By default, log data is stored in CloudWatch Logs indefinitely." Set a retention period for each log group. Log groups and streams