Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 2

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.

Question 11

An invoice assistant must show an amount and a traceable reference to the exact source line that justifies it. The model currently invents totals when scanned invoices have missing lines. What design fixes the audit requirement?

Ask the model to state that it is confident about the number, and retain only the generated response.
Increase retrieved chunk count without retaining which invoice page and line supplied the amount.
Use a higher-temperature generation setting to explore several totals before selecting one.
Preserve source-page/field provenance and reconcile amounts with deterministic extraction and calculation before the FM explains them.

Explanation

Correct Answer

Source spans identify the invoice evidence and deterministic checks validate amounts independently; the generated explanation cites those verified facts.

Incorrect Answers

A confidence statement generated by a model is not an audit trail or evidence that the amount is correct.

More retrieved text may help recall but does not provide line-level lineage or ensure arithmetic correctness.

Sampling variation is not a validated financial reconciliation method.

Question 12

Two divisions share one Bedrock model endpoint. Their records, conversation histories and permitted business actions must remain isolated. Which THREE independent boundaries must the application enforce?

Authorize the requesting identity against the tenant-owned source records.
Constrain retrieval to authorized tenant documents before placing passages in the model context.
Separate tenant session/memory state and enforce action permissions at each tool or API.
Use a common encrypted vector index without tenant filters and assume encryption enforces document-level separation.
Use a system prompt stating that the model must never mention the other division, with no data-access checks.
Give every agent tool a shared administrator role so the same deployment behaves consistently.

Explanation

Correct Answer

Tenant-scoped authorization prevents a caller from reading records solely because the service role can access them.

Filtering before model context is necessary: an answer-stage prompt cannot reliably remove private passages that were already supplied.

Memory/session isolation prevents history leaks; resource-level tool checks prevent cross-tenant writes despite model-generated action requests.

Incorrect Answers

Encryption at rest protects stored bytes but does not decide whether a tenant may retrieve a particular document.

A prompt is not an authorization engine; it cannot enforce row-level filtering or prevent inappropriate source inclusion.

A shared admin identity allows privilege confusion and does not preserve caller-specific action rights.

Question 13

A regulated service may propose a customer refund, but policy requires a named employee to approve it before funds move. Where should the approval gate be placed?

Include an instruction telling the model that it should approve only refunds it thinks are reasonable.
Accept an approval string emitted by the model as equivalent to an authenticated employee signature.
Enforce authenticated human approval in the transaction service before executing the refund tool.
Log every proposed refund after execution and assume retrospective logging replaces pre-approval.

Explanation

Correct Answer

The payment/refund boundary can verify identity, scope and approval receipt independently of the model before committing funds.

Incorrect Answers

A model recommendation is not a verified approver and cannot demonstrate that an authorized employee reviewed a particular transaction.

Generated text is not evidence of an authenticated person or a recorded approval decision.

Audit logging is important but cannot satisfy a requirement that approval occur before the financial action.

Question 14

A proof of concept answers expected test prompts correctly but has never encountered malformed inputs, a missing Knowledge Base result or a timed-out billing tool. Which acceptance-test change is most important?

Increase the size of the prompt template while keeping only successful example questions.
Add failure cases for invalid requests, missing retrieval evidence and dependency timeouts, with expected safe outcomes.
Choose a higher-ranked model and assume its training data covers these integration failures.
Measure only average token usage on the successful prompts and infer failure handling from that average.

Explanation

Correct Answer

Negative-path cases test whether the application rejects invalid inputs, abstains without evidence and reports unknown transaction status without hallucinating success.

Incorrect Answers

Additional happy-path context does not establish failure behavior for unavailable dependencies or malformed data.

Model capability does not replace application-level exception, timeout and fallback tests.

Token usage is a performance or cost signal, not proof of safe behavior when retrieval or tools fail.

Question 15

A GenAI application design contains no correlation between user requests, Knowledge Base retrieval, model inference and downstream tool calls. Which observability plan closes that gap?

Carry an application correlation ID across retrieval, model and tool stages; capture token usage, stage latency and accepted-answer quality without indiscriminate sensitive payloads.
Enable full raw prompt logging for every model endpoint and infer all downstream tool events from its contents.
Count HTTP 200 model responses and treat those as the number of correct, committed business operations.
Record only the model identifier and omit retrieval status, tool receipts and user-facing completion times.

Explanation

Correct Answer

Correlated application telemetry distinguishes stages and supports latency/cost investigation; quality metrics check business outcomes and sensitive data is minimized.

Incorrect Answers

Bedrock invocation logging is endpoint- and operation-specific; it does not by itself capture every app retrieval or actual tool commit. Raw logging may expose private data.

An HTTP-successful inference can return a wrong claim or request a tool that was never executed.

A model ID cannot explain whether slow retrieval, token generation or a timed-out tool caused the user-visible failure.

Question 16

Several teams each call supported Bedrock models with slightly different request adapters. Their applications inconsistently handle message roles, tool results and throttling. What design change most directly improves integration consistency?

Duplicate SDK wrappers in each team and let every app interpret model errors differently.
Provide a versioned, tested inference adapter using supported Converse/InvokeModel contracts and normalized error/tool handling.
Put AWS credentials into each user-visible prompt so model requests do not need an authenticated service tier.
Remove request schema validation because model output is unstructured by nature.

Explanation

Correct Answer

A controlled adapter centralizes validated request/response formats and error behavior, while checking whether each model supports the chosen API and features.

Incorrect Answers

More independent wrappers perpetuate schema and retry drift and increase the number of places where a tool-result protocol can break.

Service credentials must not be delegated through a prompt; the application should call Bedrock under scoped AWS IAM permissions.

Even when generated text varies, API message structure and any tool arguments require strong validation.

Question 17

A production assistant becomes unavailable when its primary model or Region fails. The business permits reduced functionality but requires a truthful response and tested alternate inference. What is the most direct resilience design?

Configure health-based routing to a pre-qualified alternative model/Region, verify permitted destinations, and report degraded behavior when needed.
Keep invoking the failed model until it eventually returns, without deadline or fallback.
Change the prompt to claim success after timeouts because downstream operations might later recover.
Assume any foundation model identifier in another Region accepts the same API and output contract without qualification.

Explanation

Correct Answer

A failover path must be tested for model capabilities, region/IAM permissions, quality and schema differences, and must not imply successful work that did not complete.

Incorrect Answers

Unbounded retries make an outage longer and can exhaust latency budgets and service quotas.

A timeout leaves a request uncertain; fabricated completion misinforms users and can trigger unsafe business actions.

Model and endpoint support vary by Region and version, so an untested identifier is not a reliable recovery plan.

Question 18

An architecture review reveals that agent tools can change customer records but nobody owns the approval policy or the tool's effective permissions. What governance design is required before enabling those actions?

Assign accountable service owners and enforce independently verified caller authorization for each action at the tool or business API.
Allow the foundation model to select its own IAM permissions based on the prompt at runtime.
Give one administrator credential to all tools to avoid responsibility gaps between teams.
Delay permission reviews until after production and rely solely on a daily list of completed writes.

Explanation

Correct Answer

Named owners set the policy and maintain the interface; downstream resource checks bind each requested mutation to a permitted identity and audit record.

Incorrect Answers

Model text cannot legitimately grant an IAM role new permissions or substitute for a policy decision.

Universal administrator access erases least privilege and caller accountability.

After-the-fact reports do not prevent unauthorized record changes or identify who should approve specific controls.

Question 19

Policy documents change hourly, but a demonstration assistant still searches yesterday's index. What must be verified before the system can claim to answer from the current approved policy?

Keep the existing index and ask the model to infer the missing policy changes based on its pretrained knowledge.
Use a larger retrieval top-k with no ingestion or approval-status update.
Lower temperature to zero and assume a stable answer is also a current answer.
Synchronize approved policy revisions, retain effective-date metadata and verify that obsolete versions are not returned as current.

Explanation

Correct Answer

Ingestion synchronization plus version-aware filtering controls document currency; a retrieval test confirms which policy actually grounds the response.

Incorrect Answers

Pretrained knowledge does not include an hourly change unless the application supplies that change as evidence.

More results from a stale index can amplify obsolete content rather than retrieve the new provision.

Repeatable generation is not source freshness or verification of effective dates.

Question 20

A planned GenAI application has not checked whether its selected model, inference endpoint and expected request volume are supported in its required AWS Region. What should the team validate before committing to the design?

Check exact model and runtime API availability, inference profile destinations, quotas and permissions in the deployment Region.
Assume every model offers identical modalities and capacity in all AWS Regions.
Choose a public model benchmark winner and postpone regional feasibility checks until after launch.
Use S3 cross-Region replication to guarantee that a Bedrock model exists in the target Region.

Explanation

Correct Answer

Catalog/Region, endpoint capability, IAM/SCP and service quota verification establishes that the proposed runtime can actually serve the business workload.

Incorrect Answers

Foundation model support differs by model version, API and Region; a universal-availability assumption is unsafe.

Offline benchmark quality cannot make an unavailable model callable or provide required throughput.

Replicating source objects does not provision or authorize a foundation model in another inference Region.

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

Question 10

Explanation

Correct Answer

Incorrect Answers

Question 11

Explanation

Correct Answer

Incorrect Answers

Question 12

Explanation

Correct Answer

Incorrect Answers

Question 13

Explanation

Correct Answer

Incorrect Answers

Question 14

Explanation

Correct Answer

Incorrect Answers

Question 15

Explanation

Correct Answer

Incorrect Answers

Question 16

Explanation

Correct Answer

Incorrect Answers

Question 17

Explanation

Correct Answer

Incorrect Answers

Question 18

Explanation

Correct Answer

Incorrect Answers

Question 19

Explanation

Correct Answer

Incorrect Answers

Question 20

Explanation

Correct Answer

Incorrect Answers

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