Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 1

Question 1

A contact-center assistant must answer policy questions with citations and may also update a customer case, but only when the signed-in agent is authorized. Which architecture meets both needs without treating model text as permission to change records?

Use RetrieveAndGenerate for both policy answers and direct database updates; the model response itself is the update authorization.
Store policies in an S3 bucket and embed the full case-update API token in the prompt so the model can perform writes.
Use a single text-only prompt for search and mutations, then accept every well-formed output as an approved action.
Ground policy answers in permitted Knowledge Base documents and expose case mutations as separately authorized, validated application tools.

Explanation

Correct Answer

Knowledge Base retrieval supplies cited policy context; a tool/API layer independently checks identity, record-level permission, arguments and transaction outcome for each mutation.

Incorrect Answers

RetrieveAndGenerate retrieves evidence and creates text with citations; it does not grant business authorization or execute a case-record transaction.

A credential in a prompt can be disclosed and bypasses caller-specific permission checks. S3 storage alone is not an authorized tool boundary.

Well-formed model text can be incorrect or adversarial. The application must validate caller rights and action parameters before executing a write.

Question 2

A regulated EU service must ensure customer prompts are processed only inside an approved EU geography. Its S3 source documents already reside in an EU Region. What should the architect require before choosing the Bedrock inference path?

Use any global cross-Region inference profile because the S3 bucket is in the EU.
Check the candidate model and inference profile destination Regions, select an EU-compliant route, and permit only those Regions through IAM and SCPs.
Set the AWS SDK caller Region to an EU Region and ignore where the chosen inference profile can route computation.
Encrypt prompts before submission and assume encryption automatically guarantees EU-only processing.

Explanation

Correct Answer

Geographic profiles have defined destination Regions. Validate the exact model/profile and associated authorization policies so prompt processing remains in the allowed geography.

Incorrect Answers

Global profiles are not an EU geographic boundary; locating the source bucket in the EU does not constrain remote model processing destinations.

The source Region of an API request can differ from the inference execution Region when a profile routes it. Caller location alone is insufficient.

Encryption protects data in transit, but it does not restrict the AWS Region that performs inference or determine an approved processing jurisdiction.

Question 3

A claims-processing GenAI pilot has no verified outcome labels. Before choosing a foundation model, which TWO prerequisites are needed to assess claims decisions against business-quality requirements?

Create a labeled reference set of claims outcomes, with a documented verification method.
Define measurable acceptance criteria for decision accuracy, unsafe errors and escalation against that reference set.
Choose the model with the largest public benchmark score and treat that score as the claims acceptance result.
Launch an unreviewed trial that counts the number of model responses but never checks claims outcomes.
Infer approval quality from prompt token counts rather than comparing results with ground truth.

Explanation

Correct Answer

A case-level verified reference establishes what the correct claims decision actually is; without labels, apparent fluency cannot be scored for correctness.

Acceptance thresholds convert a business objective into a testable gate, including costly false approvals and cases that must be escalated.

Incorrect Answers

A general model benchmark may not represent regulated claims evidence or the failure costs of this workload.

A response count measures volume, not whether a claims outcome is correct or safe; an unlabeled trial cannot validate the decision process.

Token consumption measures usage or cost, not the truth of the claims decision. Ground-truth comparison is required.

Question 4

A policy assistant produces plausible answers but cannot identify which approved document supports a sentence. The application must also respect a caller's document access. Which change addresses the missing evidence?

Lower model temperature to zero so every generated sentence is automatically a documented citation.
Fine-tune the model on a copy of the policies and assume the model can return a valid document URL for each fact.
Use permission-filtered document retrieval and retain source identifiers/citations when generating an answer.
Give the model all confidential documents and ask in the system prompt to hide those that the caller cannot view.

Explanation

Correct Answer

Knowledge Bases can retrieve relevant source chunks and RetrieveAndGenerate can return citations; caller-specific authorization must also be enforced on retrieved material.

Incorrect Answers

Temperature affects sampling, not provenance. Deterministic output can still invent a source or cite an inaccessible document.

Changing model weights is not a citation mechanism. The app needs current source chunks and explicit attribution to the approved document.

Prompt instructions do not prevent unauthorized content from entering model context. Permissions belong in retrieval and data-access enforcement.

Question 5

A regulatory team updates its policy manuals every day. A developer proposes retraining a foundation model daily so a compliance assistant answers from current rules. Which approach is better for the rapidly changing factual content?

Increase the model's output token limit so it can infer the most recent rules without retrieving them.
Keep a frozen copy of the policy manuals in the prompt and rely on the model to infer later revisions.
Change the model temperature to zero and skip any synchronization of the authoritative source.
Synchronize approved policies into a Knowledge Base/RAG index and retrieve version-valid content when answering.

Explanation

Correct Answer

RAG separates model behavior from evolving facts. Keep the source/index synchronized and filter to the effective approved policy version during retrieval.

Incorrect Answers

A larger output budget permits more text, not access to regulations that changed after model training.

A static prompt snapshot becomes obsolete; the model cannot reliably reconstruct a newly effective policy not supplied to it.

Sampling settings do not update external evidence or resolve changed effective dates.

Question 6

A finance assistant must compute auditable account totals and then explain them conversationally. The model's arithmetic occasionally disagrees with the ledger. Which architecture is appropriate?

Increase the model temperature so alternative calculation steps can converge on a better result.
Ask the model to generate its own ledger values and treat fluent explanations as the source of truth.
Summarize the ledger first and use the summary text alone as the numerical reconciliation evidence.
Compute totals in a validated deterministic service, then provide the verified result to the FM solely for narrative explanation.

Explanation

Correct Answer

Separating execution from prose makes arithmetic testable. A tool can return the checked total, and the model can explain that verified result without recomputing it.

Incorrect Answers

Increasing randomness cannot enforce numerical correctness and may make a ledger calculation less repeatable.

A model-generated amount is not a posted or reconciled financial record; the ledger or verified calculation service must control the number.

Summarization can omit rows and change magnitudes. A financial total requires deterministic aggregation and audit linkage to source records.

Question 7

A GenAI proof of concept scores well for answer quality, but users report slow replies at production-like load. Which TWO checks directly determine whether the service meets its documented p95 end-to-end latency SLO?

Measure only the number of completed requests and omit response times for slow or failed requests.
Take a single idle-user time-to-first-token measurement and use it as final answer completion time.
Measure p95 end-to-end completion latency while running the representative concurrency and request mix.
Compare that measured p95 against the agreed production latency SLO, including timeout/failure handling.
Replace the evaluation dataset with much shorter prompts solely to make observed latency fall.

Explanation

Correct Answer

Concurrent representative traffic reveals queuing and contention that are invisible in isolated fast demos.

A documented service-level objective gives a measurable pass/fail threshold for the actual end-to-end latency distribution.

Incorrect Answers

Throughput alone omits the distribution of response times and can hide extreme tail delays.

Time to first token excludes generation, retrieval and tool execution needed for the usable final answer.

Shortening workload inputs invalidates a performance comparison unless it represents the business request distribution.

Question 8

Remote factories must continue processing inspection reports even when the WAN has been disconnected for hours. A proposed design calls a managed Bedrock API for every report. What should be decided first?

Choose a geographic cross-Region profile and assume it works without connectivity to an AWS API endpoint.
Schedule all offline requests for a Bedrock batch job and expect results during the disconnection.
Assess an edge-compatible local inference path and its hardware/synchronization needs rather than making cloud API calls mandatory offline.
Increase the Bedrock request timeout so packets can traverse a disconnected network.

Explanation

Correct Answer

Offline service requires computation and dependencies at the factory, with later synchronization when connectivity returns.

Incorrect Answers

A cross-Region profile can route AWS inference, but it still requires network access to the service.

Bedrock batch inference runs in the cloud and returns later through AWS storage; it cannot supply responses during a complete WAN outage.

A timeout cannot restore physical connectivity or make a cloud-hosted service available without a network route.

Question 9

A pilot team claims a GenAI assistant delivers positive ROI because its replies sound human. What measurement design is necessary before approving its business case?

Use a public fluency benchmark as a substitute for business outcomes and costs.
Measure accepted work saved, correction and human-escalation costs, and actual inference spending against a baseline.
Count all model-generated words as time saved even when operators must redo the answer.
Use the model parameter count as a proxy for operational savings.

Explanation

Correct Answer

ROI depends on measurable changes in labor, accepted task completion, error remediation, escalation and billed inference costs.

Incorrect Answers

Language quality does not quantify productive outcomes or the cost of wrong suggestions in this business workflow.

Length is not value; an incorrect answer may increase handling time and require a more expensive correction.

Parameter count is a model characteristic, not a measurement of savings or cost per successful case.

Question 10

A team wants to send thousands of nightly document summaries through the same synchronous path used by its interactive chatbot. Summaries are needed next morning, not during a conversation. Which approach fits these deadlines?

Send every summary in a synchronous user session and keep the connection open for the full nightly workload.
Use tool calling in a Bedrock batch record to maintain a multi-turn transaction for every document.
Separate interactive inference from a supported asynchronous Bedrock batch job using S3 inputs and completed outputs.
Use model context-window size as the only scheduling control and ignore job completion deadlines.

Explanation

Correct Answer

Asynchronous batch is designed for independent record processing with an S3 input and S3 output after completion; interactive callers keep a responsive path.

Incorrect Answers

Keeping user sessions open for bulk jobs needlessly couples interactive availability to long-running processing.

Batch inference does not support back-and-forth tool calling inside a record. Interactive tool workflows must use a supported real-time path.

Context capacity determines request size, not asynchronous scheduling, completion monitoring or end-user responsiveness.

Pass Your Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Exam Easy!

Question 1

Explanation

Correct Answer

Incorrect Answers

Question 2

Explanation

Correct Answer

Incorrect Answers

Question 3

Explanation

Correct Answer

Incorrect Answers

Question 4

Explanation

Correct Answer

Incorrect Answers

Question 5

Explanation

Correct Answer

Incorrect Answers

Question 6

Explanation

Correct Answer

Incorrect Answers

Question 7

Explanation

Correct Answer

Incorrect Answers

Question 8

Explanation

Correct Answer

Incorrect Answers

Question 9

Explanation

Correct Answer

Incorrect Answers

Question 10

Explanation

Correct Answer

Incorrect Answers

AWS Certified Generative AI Developer - Professional AIP-C01 Practice Tests