CLOUDPUZ
AWS Certified
Generative AI Developer -
Professional
AIP-C01
150 Practice Questions
Answers & Detailed Explanations
2 practice tests • 75 questions per test • English
Independent study material for certification preparation. This publication is not affiliated with or endorsed by Amazon Web
Services. AWS and related names are trademarks of their respective owners.
Practice questions for learning and revision; no claim is made that these are live certification exam questions.
CLOUDPUZ | AIP-C01 | Practice questions & answers 1
, How to use this question bank
Complete each practice test before consulting its answer key. Select one answer unless the
question specifies multiple answers. Review the explanations for every option, then revisit the
topics you missed.
Each test has three sections: Questions, Quick answer key, and Answers with explanations.
Question numbering restarts at 1 for Test 2.
The questions and explanations retain the content of the Cloudpuz source question banks. Official reference links supplied
in those banks are preserved for further study. Service capabilities and exam objectives may evolve.
CLOUDPUZ | AIP-C01 | Practice questions & answers 2
, Practice Test 1
Questions 1-75
Question 1 | D5
A team changes a Bedrock prompt to request clearer explanations. Monday's baseline scores
were generated before a knowledge-base update, while Friday's candidate scores used the
expanded corpus and a different model revision. The candidate scores are higher. The team
needs evidence of the prompt change's effect on its existing benchmark, independently of the
other changes. It can replay the benchmark outside production and has retained the original
retrieved contexts. Which experiment best supports that conclusion?
A. Raise the candidate's output limit and temperature until its judge score exceeds the
archived baseline, then record the candidate as the winner.
B. Compare both prompts on identical held-out questions and saved contexts, fixing the
model, inference settings, and evaluation rubric.
C. Compare Monday's baseline mean with Friday's candidate mean again using a larger set of
newly collected customer questions.
D. Use the candidate prompt only on longer questions and the baseline only on shorter
questions, then compare their average judge scores.
Select 1 answer.
Question 2 | D3
A construction support agent retrieves supplier bulletins through a search tool. One bulletin
includes text claiming to be a new system instruction and asks the assistant to reveal its internal
operating prompt. The orchestrator currently appends every search result to its system-message
string. User-input prompt-attack filtering is enabled, but this text arrived inside a tool result. The
company must continue to use supplier material as evidence. Which change most directly fixes
the trust-boundary error?
A. Supply bulletins as tool-result evidence, preserve trusted instructions separately, and treat
document instructions as data.
B. Retrieve fewer bulletins and allow the model to decide whether each returned document
should replace the current system instructions.
C. Allow only bulletins whose text begins with the supplier company name and keep the
existing system-message assembly process.
D. Increase the prompt-attack strength and continue merging every retrieved bulletin into the
system-message string.
Select 1 answer.
CLOUDPUZ | AIP-C01 | Practice questions & answers 3
, Question 3 | D4
An interactive assistant uses Lambda workers to invoke Amazon Bedrock. A downstream limiter
already caps simultaneous model calls at a tested safe level. During a load test, model service
time stays stable, but p95 time waiting for a limiter permit rises from 100 ms to 7 seconds. The
company cannot increase downstream capacity for this release. Clients support explicit overload
responses with delayed retries. The goal is to protect useful responses from excessive waiting
and avoid spending inference tokens on abandoned work. What should the developer change?
A. Buffer excess requests in SQS and process the backlog in arrival order without checking
their original request deadlines.
B. Increase every request timeout until the longest observed queue wait can be
accommodated.
C. Increase Lambda concurrency so more workers can wait for the same fixed number of
downstream permits.
D. Bound the waiting queue, enforce request deadlines before inference, and return explicit
retryable overload responses when admission capacity is exhausted.
Select 1 answer.
Question 4 | D4
Three departments share one Lambda inference gateway and the same on-demand Bedrock
model. Finance wants AWS billing reports to distinguish each department's model charges, rather
than dividing a combined bill by request count. Requests have very different token lengths. The
gateway already knows the authenticated department, but all calls currently use the foundation
model ID. Which two changes establish native department-level cost attribution for future calls
without duplicating the gateway or purchasing separate provisioned capacity? Select TWO.
A. Tag the shared Lambda function with all three department names and use those tags to
divide Bedrock model charges automatically.
B. Configure API Gateway usage plans with an equal monthly request allowance for each
department and allocate one-third of the model bill to each plan.
C. Create an application inference profile for each department, attach a department
cost-allocation tag, and activate the relevant user-defined tags for billing.
D. Choose the authenticated department's application inference profile ARN as modelId when
the gateway invokes Bedrock.
E. Record the department in the prompt text and enable CloudTrail so AWS Billing can allocate
inference charges from that text.
Select 2 answers.
CLOUDPUZ | AIP-C01 | Practice questions & answers 4
AWS Certified
Generative AI Developer -
Professional
AIP-C01
150 Practice Questions
Answers & Detailed Explanations
2 practice tests • 75 questions per test • English
Independent study material for certification preparation. This publication is not affiliated with or endorsed by Amazon Web
Services. AWS and related names are trademarks of their respective owners.
Practice questions for learning and revision; no claim is made that these are live certification exam questions.
CLOUDPUZ | AIP-C01 | Practice questions & answers 1
, How to use this question bank
Complete each practice test before consulting its answer key. Select one answer unless the
question specifies multiple answers. Review the explanations for every option, then revisit the
topics you missed.
Each test has three sections: Questions, Quick answer key, and Answers with explanations.
Question numbering restarts at 1 for Test 2.
The questions and explanations retain the content of the Cloudpuz source question banks. Official reference links supplied
in those banks are preserved for further study. Service capabilities and exam objectives may evolve.
CLOUDPUZ | AIP-C01 | Practice questions & answers 2
, Practice Test 1
Questions 1-75
Question 1 | D5
A team changes a Bedrock prompt to request clearer explanations. Monday's baseline scores
were generated before a knowledge-base update, while Friday's candidate scores used the
expanded corpus and a different model revision. The candidate scores are higher. The team
needs evidence of the prompt change's effect on its existing benchmark, independently of the
other changes. It can replay the benchmark outside production and has retained the original
retrieved contexts. Which experiment best supports that conclusion?
A. Raise the candidate's output limit and temperature until its judge score exceeds the
archived baseline, then record the candidate as the winner.
B. Compare both prompts on identical held-out questions and saved contexts, fixing the
model, inference settings, and evaluation rubric.
C. Compare Monday's baseline mean with Friday's candidate mean again using a larger set of
newly collected customer questions.
D. Use the candidate prompt only on longer questions and the baseline only on shorter
questions, then compare their average judge scores.
Select 1 answer.
Question 2 | D3
A construction support agent retrieves supplier bulletins through a search tool. One bulletin
includes text claiming to be a new system instruction and asks the assistant to reveal its internal
operating prompt. The orchestrator currently appends every search result to its system-message
string. User-input prompt-attack filtering is enabled, but this text arrived inside a tool result. The
company must continue to use supplier material as evidence. Which change most directly fixes
the trust-boundary error?
A. Supply bulletins as tool-result evidence, preserve trusted instructions separately, and treat
document instructions as data.
B. Retrieve fewer bulletins and allow the model to decide whether each returned document
should replace the current system instructions.
C. Allow only bulletins whose text begins with the supplier company name and keep the
existing system-message assembly process.
D. Increase the prompt-attack strength and continue merging every retrieved bulletin into the
system-message string.
Select 1 answer.
CLOUDPUZ | AIP-C01 | Practice questions & answers 3
, Question 3 | D4
An interactive assistant uses Lambda workers to invoke Amazon Bedrock. A downstream limiter
already caps simultaneous model calls at a tested safe level. During a load test, model service
time stays stable, but p95 time waiting for a limiter permit rises from 100 ms to 7 seconds. The
company cannot increase downstream capacity for this release. Clients support explicit overload
responses with delayed retries. The goal is to protect useful responses from excessive waiting
and avoid spending inference tokens on abandoned work. What should the developer change?
A. Buffer excess requests in SQS and process the backlog in arrival order without checking
their original request deadlines.
B. Increase every request timeout until the longest observed queue wait can be
accommodated.
C. Increase Lambda concurrency so more workers can wait for the same fixed number of
downstream permits.
D. Bound the waiting queue, enforce request deadlines before inference, and return explicit
retryable overload responses when admission capacity is exhausted.
Select 1 answer.
Question 4 | D4
Three departments share one Lambda inference gateway and the same on-demand Bedrock
model. Finance wants AWS billing reports to distinguish each department's model charges, rather
than dividing a combined bill by request count. Requests have very different token lengths. The
gateway already knows the authenticated department, but all calls currently use the foundation
model ID. Which two changes establish native department-level cost attribution for future calls
without duplicating the gateway or purchasing separate provisioned capacity? Select TWO.
A. Tag the shared Lambda function with all three department names and use those tags to
divide Bedrock model charges automatically.
B. Configure API Gateway usage plans with an equal monthly request allowance for each
department and allocate one-third of the model bill to each plan.
C. Create an application inference profile for each department, attach a department
cost-allocation tag, and activate the relevant user-defined tags for billing.
D. Choose the authenticated department's application inference profile ARN as modelId when
the gateway invokes Bedrock.
E. Record the department in the prompt text and enable CloudTrail so AWS Billing can allocate
inference charges from that text.
Select 2 answers.
CLOUDPUZ | AIP-C01 | Practice questions & answers 4
