AIP-C01
Free AIP-C01 practice questions
10 questions from our AIP-C01 practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Foundation Model Integration, Data Management, and Compliance
A media company's content-tagging Lambda functions call Amazon Bedrock with a hard-coded model ID. The ML team evaluates new models every few weeks and wants to switch the model ID, inference parameters, and fallback model per business unit without redeploying function code. Changes must roll out gradually and roll back automatically if the functions' error-rate alarm fires. Which approach meets these requirements?
- AStore the model settings in Lambda environment variables and change them through an AWS CloudFormation stack update for each model change.
- BStore the settings as SecureString parameters in AWS Systems Manager Parameter Store and read them on every invocation of the functions.
- CKeep the settings in an AWS AppConfig profile read via the Lambda extension, deployed gradually with the alarm as a rollback trigger. Correct answer
- DPublish a new Lambda version for each model change and shift traffic with a weighted alias managed by AWS CodeDeploy.
Answer and explanation: AppConfig separates configuration from code, validates it, and rolls changes out with deployment strategies that automatically roll back when an associated CloudWatch alarm goes into ALARM, while the Lambda extension caches the values locally. Environment-variable updates through CloudFormation change the function configuration all at once with no alarm-driven gradual rollout. Parameter Store has no deployment strategies or automatic rollback, and per-invocation reads add latency. CodeDeploy alias shifting is gradual but requires publishing new function versions for every model change.
Question 2 of 10, Foundation Model Integration, Data Management, and Compliance
A marketing team uses Amazon Bedrock Flows to generate localized promotional emails. Before the prompt runs, product SKUs in the input must be resolved to current prices through an internal pricing API. After generation, any email longer than 900 characters or missing an unsubscribe line must be rejected. The team wants both steps inside the same flow. Which flow design meets these requirements?
- AReplace the flow with a Bedrock agent whose action group performs the pricing lookup, and instruct the agent to keep every email under 900 characters.
- BAdd a Lambda node before the prompt to enrich SKUs from the pricing API and one after it to validate the email, with a condition node routing failures to rejection. Correct answer
- CPut the pricing lookup instructions and the validation rules in the prompt node text, and let the model call the internal pricing API and check its own email before returning it.
- DAttach a guardrail to the prompt node with a word filter for the unsubscribe line and a maximum output length so that noncompliant emails are blocked.
Answer and explanation: Lambda function nodes run custom code inside a flow, so one can enrich the input from the pricing API before generation and another can deterministically validate length and the unsubscribe line, while a condition node routes failures. A prompt node cannot call external APIs, and model self-checking is not deterministic. Guardrail word filters block listed terms; they cannot require text to be present or enforce length. An agent adds autonomous planning where a fixed pipeline is wanted and provides no post-generation validation step.
Question 3 of 10, Foundation Model Integration, Data Management, and Compliance
A regulated insurer manages prompts in Amazon Bedrock Prompt Management. Policy states that prompt authors may edit drafts, but only a compliance group may create the prompt versions that production uses, production may run only those approved versions, and every change must remain auditable for seven years. Which solution meets these requirements?
- ATag approved prompt versions with approved=true and have the production code check the tag before each invocation.
- BLet authors UpdatePrompt but only compliance CreatePromptVersion, pin production to version ARNs, and keep CloudTrail in Object Lock S3 for 7 years. Correct answer
- CKeep prompts in a Git repository with mandatory pull-request approval, then paste the merged text into the prompt draft that production calls.
- DAllow all authors to create versions and have the compliance group review a CloudWatch dashboard weekly to revert any unapproved versions.
Answer and explanation: Separating draft updates from version creation in IAM makes the compliance role the only publisher of immutable versions, production pins those version ARNs, and CloudTrail logs in an Object Lock bucket provide tamper-resistant seven-year evidence. Letting everyone publish and reviewing weekly is detective rather than preventive. Pasting approved text into the draft means production runs a mutable draft that authors can still edit. A tag check is an honor system that authors could change themselves.
Question 4 of 10, Implementation and Integration
A developer-tools company is adding an AI assistant to its web IDE. Responses from Amazon Bedrock can take more than 60 seconds for long code explanations, and product managers require the first tokens to appear in the browser in under two seconds. The team wants the fewest moving parts and already uses AWS Lambda for its backend logic. Which approach should the team implement?
- AAmazon API Gateway WebSocket API whose Lambda function calls the Converse API and then sends the full answer to the client with a single PostToConnection call
- BA Lambda function URL configured with the BUFFERED invoke mode whose handler calls the ConverseStream API and concatenates the deltas before returning
- CAmazon API Gateway HTTP API with a Lambda proxy integration that calls the Converse API and returns the complete answer as JSON
- DA Lambda function URL configured with the RESPONSE_STREAM invoke mode whose handler calls the ConverseStream API and writes each content delta to the response stream Correct answer
Answer and explanation: A function URL in RESPONSE_STREAM mode lets the handler forward ConverseStream deltas to the browser as they arrive, giving fast time to first token and supporting responses longer than typical API integration timeouts with minimal infrastructure. An HTTP API proxy integration returns only after the full answer is ready and is bounded by the integration timeout. Sending the whole answer with one PostToConnection call wastes the WebSocket's ability to push incremental chunks. BUFFERED mode holds the entire payload until the function finishes, so streaming from Bedrock never reaches the user.
Question 5 of 10, Implementation and Integration
An analytics company lets users request a 40-page market report that Amazon Bedrock generates in three to five minutes through several chained prompts. Users submit requests from mobile apps on unreliable networks, and a dropped connection must never lose a finished report. The company wants a managed, serverless design. Which approach meets these requirements?
- AStream the report to the app through a Lambda function URL in RESPONSE_STREAM mode so that the user sees progress immediately as sections are generated
- BRequest an increase of the API Gateway integration timeout and keep the HTTP connection open until the final prompt in the chain completes
- CSubmit each request as an Amazon Bedrock batch inference job, return the job ARN to the mobile app, and have the app poll for the job status
- DReturn 202 with a job ID, run the prompt chain in a Step Functions Standard workflow that saves the report to S3 and status to DynamoDB, and push-notify the app Correct answer
Answer and explanation: The asynchronous request-reply pattern decouples generation from the client connection: the workflow persists the result and status, so a dropped network only delays retrieval. Holding a long HTTP connection open, even with a higher timeout, loses the response when the mobile network drops. Streaming improves perceived latency but still loses the output if the connection breaks mid-stream. Batch inference is designed for large offline datasets and does not offer interactive turnaround for individual requests.
Question 6 of 10, Implementation and Integration
A telecom provider receives free-text support requests that fall into billing disputes, network outages, and device troubleshooting. Each category is best handled by a different Amazon Bedrock model with its own prompt. Compliance requires a visual, auditable record of which path each request took, and each model call needs its own retry policy. Which solution meets these requirements?
- AAn Amazon Bedrock intelligent prompt router configured with the three models so that each request is routed to the best model automatically
- BA Step Functions Standard workflow that classifies with a small model, routes with a Choice state to a category-specific Bedrock task, and sets Retry per task Correct answer
- CAn Amazon SNS topic with message filtering policies that deliver requests to three subscriber Lambda functions based on a category attribute set by the client
- DA single Lambda function that classifies the request with a small model and then calls the category-specific model in an if-else block with SDK retries
Answer and explanation: Step Functions records every state transition in the execution history, provides a visual path for audits, and lets each Bedrock task define its own Retry configuration, while the Choice state performs content-based routing. A single Lambda function works but leaves no visual per-request audit trail and hides routing inside code. Intelligent prompt routing chooses between models in the same family to balance quality and cost, not between domain-specialized prompts by category. SNS filtering depends on the client already knowing the category, which is the problem to be solved.
Question 7 of 10, AI Safety, Security, and Governance
A telecom company generates call summaries with Amazon Bedrock. Summaries must still be produced when transcripts contain customer email addresses and phone numbers, but those values must never appear in the stored summary. If an agent pastes a full payment card number into the prompt, the request must be rejected outright. Which Amazon Bedrock Guardrails configuration meets these requirements?
- AConfigure the sensitive information filter with the EMAIL and PHONE entity types set to mask and the credit or debit card number entity type set to block. Correct answer
- BConfigure the sensitive information filter with all three entity types set to mask, and add a contextual grounding check to reject summaries containing card numbers.
- CAdd a word filter containing regular expressions for email addresses, phone numbers, and card numbers with the action set to block.
- DConfigure a denied topic named payment data with sample phrases, and enable the HIGH strength misconduct content filter for outputs.
Answer and explanation: Sensitive information filters let each PII entity type have its own action, so masking email and phone values keeps summaries flowing while blocking card numbers rejects those requests. Word filters match exact words and phrases rather than patterns, and blocking email and phone would stop summaries that must still be produced. A denied topic and content filter are semantic classifiers that do not reliably detect or redact specific PII values. Masking card numbers would let the request proceed, and grounding checks measure faithfulness to sources rather than detecting PII.
Question 8 of 10, AI Safety, Security, and Governance
A retail bank is launching a public chatbot on Amazon Bedrock. Marketing requires that the bot never mention four named competitor banks, matching their exact brand names. Legal requires that the bot decline any request for personalized investment recommendations, however the user phrases it. The bank wants to meet both requirements with managed Amazon Bedrock Guardrails policies. Which configuration should the developer use?
- ADefine a denied topic for each competitor bank and add the phrase 'investment recommendation' to a custom word filter.
- BAdd the competitor brand names to a custom word filter and define a denied topic that describes personalized investment recommendations with sample phrases. Correct answer
- CAdd both the competitor names and investment-related keywords to a custom word filter and set the prompt attack filter to HIGH.
- DEnable the misconduct content filter at HIGH strength and add the competitor brand names to the sensitive information filter as custom regular expressions with the block action.
Answer and explanation: Word filters block exact words and phrases, which suits a fixed list of brand names, while denied topics use natural-language definitions to catch a subject regardless of phrasing. Using denied topics for brand names and a single keyword for investment advice reverses the strengths of each policy and misses paraphrased advice requests. The misconduct filter does not cover investment advice, and sensitive information filters are intended for PII-like patterns. Keyword lists cannot reliably catch every phrasing of advice, and the prompt attack filter targets jailbreaks, not topics.
Question 9 of 10, Operational Efficiency and Optimization for GenAI Applications
A SaaS vendor serves 40 enterprise tenants from one AWS account and calls the same Amazon Bedrock foundation model for all of them. Finance must see Bedrock inference cost per tenant directly in AWS Cost Explorer and in the Cost and Usage Report, and the platform team wants the least ongoing engineering effort. Which solution meets these requirements?
- APublish a custom CloudWatch metric for token usage with a TenantId dimension and build a CloudWatch dashboard that finance can use to estimate each tenant's share of the bill
- BMove each tenant into its own AWS account in AWS Organizations and use consolidated billing to separate the Bedrock charges by linked account
- CEnable model invocation logging, add the tenant ID to the request metadata, and run a nightly Athena query that multiplies each tenant's token counts by the published price
- DCreate a tagged application inference profile per tenant, activate the tenant tag for cost allocation, and invoke through each tenant's profile Correct answer
Answer and explanation: Application inference profiles can carry tags, and once the tenant tag is activated as a cost allocation tag, Bedrock usage invoked through each profile appears per tenant in Cost Explorer and the Cost and Usage Report. An Athena calculation works but is custom code outside the billing tools and must track price changes. A custom metric dashboard shows tokens, not billed cost in Cost Explorer. One account per tenant isolates billing but carries far more operational overhead than tagging.
Question 10 of 10, Testing, Validation, and Troubleshooting
A SaaS vendor wants users to rate each assistant answer with thumbs up or thumbs down and an optional comment. Low-rated answers must be reviewed by internal linguists who label the failure type, and every rating must be traceable to the exact prompt and response that Amazon Bedrock processed. Which design meets these requirements?
- AStore ratings in DynamoDB through an API Gateway and Lambda endpoint, keyed by the request ID in the invocation logs, and route low-rated items to a Ground Truth job with a private linguist workforce. Correct answer
- BCapture ratings as Amazon CloudWatch RUM custom page events and have linguists search the RUM console for low-rated sessions to record a failure type for each one.
- CAsk the model to self-assess every answer with a confidence score and treat answers below 0.5 as thumbs-down feedback for the linguists.
- DStore ratings and comments in browser local storage and have a nightly script upload them to Amazon S3, where the linguists download low-rated items and label them in a spreadsheet.
Answer and explanation: A feedback API that stores ratings under the same request ID used in invocation logs makes each rating traceable to the logged prompt and response, and Ground Truth with a private workforce provides a managed annotation workflow for failure labels. Local storage uploads are unreliable and lack correlation to invocations. RUM captures front-end telemetry rather than labeled review workflows, and model self-assessment is not user feedback.