Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 50

Question 490

A prompt variant generates different field labels. Which diagnostic distinguishes the likely service failure?

Assume every versioned prompt generates identical key names because its purpose is unchanged.
Compare template/model versions and validate structured schema across releases.
Accept any JSON field name that sounds similar and silently write it into the old schema.
Change only the UI label without testing the API's actual output fields.

Explanation

Correct Answer

Compare saved prompt versions, model/request settings and rendered field mappings; validate structured output contract across the new variant.

Incorrect Answers

Template changes can alter field labels and response shape.

Ambiguous field mapping can corrupt business data.

A display change does not repair a machine-readable contract regression.

Question 491

A model answers from the wrong conversation. Which troubleshooting step should be performed first?

Allow every user to reuse the same session ID and ask the model to separate their replies.
Treat a text display name as proof that the retrieved memory belongs to the caller.
Check session ID, user actor scoping and persisted memory isolation.
Clear only the browser chat history while keeping a shared backend memory namespace.

Explanation

Correct Answer

Validate user actor, AgentCore memory/session namespace and effective session ID; cross-conversation results indicate a state isolation problem.

Incorrect Answers

Session reuse risks cross-user context contamination.

Names can be spoofed and are not access credentials.

UI state clearing does not enforce storage isolation.

Question 492

A response stream ends before completion. What specific check can confirm the actual failure mechanism?

Detect stream failure and mark incomplete, not return a false successful answer.
Display the partial response as verified complete when the network connection closes unexpectedly.
Generate a plausible ending locally and label it the missing original tokens.
Ignore cancellation/error frames and always trigger the downstream transaction success callback.

Explanation

Correct Answer

Detect stream termination/transport failure and maintain an incomplete status until an actual successful stop event and output validation.

Incorrect Answers

Connection closure can truncate output.

A guessed ending is not from the invoked model stream.

Partial generation cannot authorize or confirm side effects.

Question 493

Model invocation logging is enabled for a team in us-east-1, but a second application uses a runtime endpoint in eu-west-1 and its calls do not appear in the us-east-1 log destination. What investigation identifies the cause before calling the model service broken?

Assume bedrock-mantle inference appears in the same bedrock-runtime invocation logs.
Check the actual invocation endpoint, caller account, destination Region and supported operation, then inspect the logging configuration for the Region receiving those runtime requests.
Change only IAM model permissions while leaving invocation logging disabled.
Expect 100 KB-plus multimodal payload bytes inline in every CloudWatch event.

Explanation

Correct Answer

Bedrock model invocation logging is configured by account and Region. A runtime request in another Region will not automatically appear in the first Region’s CloudWatch/S3 destination; verify its actual endpoint and enabled service logging scope.

Incorrect Answers

AWS documents endpoint-specific logging coverage.

Invocation authorization is not the logging configuration.

Large/binary payloads can be delivered separately to S3.

Question 494

A guardrail rejects valid legal-policy questions. Which troubleshooting step should be performed first?

Disable the guardrail entirely after one legitimate legal question is blocked.
Reproduce with versioned denied-topic/word filter settings and evaluate false positives.
Assume denied topics use an ordinary substring regex and fix by shortening one keyword.
Record only the model ID while ignoring the guardrail/policy version used for the failure.

Explanation

Correct Answer

Reproduce the false-positive on the exact guardrail version and denied-topic/word policy, comparing a labeled allow/deny suite before a controlled change.

Incorrect Answers

Removing moderation forfeits protection against prohibited content.

Topic classifiers and word filters behave differently and need tested scope.

The same model can be evaluated differently after a guardrail update.

Question 495

A grounded answer includes a citation to unauthorized internal content. What specific check can confirm the actual failure mechanism?

Diagnose permission filtering before retrieval rather than simply stripping citations.
Strip only the link URL while leaving the confidential text in the model's context.
Trust a generated 'safe source' label without validating the actual document ACL.
Filter only the final answer's profanity/PII and skip corpus tenant permissions.

Explanation

Correct Answer

Enforce authorization at retrieval time before a protected chunk reaches the model; deleting the visible citation after generation is too late.

Incorrect Answers

The data leak already happened during retrieval.

Model output cannot authorize a document.

Content moderation is not source ownership enforcement.

Question 496

A model update triggers off-topic responses after prompt variable expansion. Which diagnostic distinguishes the likely service failure?

Read only the unrendered prompt template and assume user variable substitutions cannot affect meaning.
Change the embedding vector index because the assistant went off topic after prompt expansion.
Check rendered prompt template, variable escaping and message role structure.
Assume any text prepended with SYSTEM inside a user variable has legitimate system authority.

Explanation

Correct Answer

Inspect actual rendered prompt variables, escaping and API role blocks across the model/prompt versions to locate unintended instruction changes.

Incorrect Answers

Injected or malformed values can change the effective prompt.

Retrieval storage is not necessarily the cause of malformed prompt roles.

Literal role headings in data cannot change trusted API message roles.

Question 497

A model output has inconsistent answer facts after retrieval. Which troubleshooting step should be performed first?

Blame the vector store without checking which correct chunks were actually passed into the generation context.
Count citations as proof all generated facts follow from them.
Only test model temperature and ignore source version or retrieval configuration changes.
Compare retrieved source chunks with generation prompts and final response independently.

Explanation

Correct Answer

Compare retrieval hits, final generation prompt and response claim-by-claim to distinguish wrong evidence from unsupported model synthesis.

Incorrect Answers

Good evidence can still be misused during generation.

A citation may not entail the surrounding claim.

Sampling cannot diagnose every retrieval-grounding failure.

Question 498

A performance incident only happens on long documents. What specific check can confirm the actual failure mechanism?

Increase answer maxTokens beyond the model's context budget without reducing input.
Change the endpoint IAM role because the error occurs only for big documents.
Analyze tokenized input size, retrieval chunk count and model context limits.
Discard the latter half of every document and present the result as complete source coverage.

Explanation

Correct Answer

Measure tokenized input length, chunk selection, parser output coverage and model context limits for long documents; the failure can arise before generation.

Incorrect Answers

The combined context/output request still exceeds limits.

Document-size-correlated errors suggest payload/context rather than permissions alone.

Silent truncation invalidates long-document claims.

Question 499

A tool timeout is labeled as successful user action. Which diagnostic distinguishes the likely service failure?

Report completed because the model originally planned to invoke the tool.
Retry a potentially non-idempotent action with a new ID without querying previous state.
Inspect downstream request IDs and persist unresolved state with truthful escalation.
Suppress the timeout error and fabricate a success receipt for the user.

Explanation

Correct Answer

Look up downstream request IDs and actual tool status; preserve pending/unknown when a timeout obscures whether the user action committed.

Incorrect Answers

Intent does not establish execution.

The first attempt may already have committed.

An invented receipt cannot prove persisted business state.

Question 500

A model change degrades multilingual recognition. Which troubleshooting step should be performed first?

Approve the new model solely because it passes an English-only benchmark.
Increase generation temperature and assume it restores missing locale support.
Remove failing language-specific test cases to retain a high overall accuracy figure.
Test version-specific locale samples and adjust routing or fallback according to evidence.

Explanation

Correct Answer

Compare version-specific multilingual labeled samples and route or rollback to a validated language-capable model when recognition regressions are confirmed.

Incorrect Answers

English scores do not cover the affected languages.

Sampling settings do not guarantee better multilingual semantics.

Excluding affected populations hides the deployment regression.

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

Question 490

Explanation

Correct Answer

Incorrect Answers

Question 491

Explanation

Correct Answer

Incorrect Answers

Question 492

Explanation

Correct Answer

Incorrect Answers

Question 493

Explanation

Correct Answer

Incorrect Answers

Question 494

Explanation

Correct Answer

Incorrect Answers

Question 495

Explanation

Correct Answer

Incorrect Answers

Question 496

Explanation

Correct Answer

Incorrect Answers

Question 497

Explanation

Correct Answer

Incorrect Answers

Question 498

Explanation

Correct Answer

Incorrect Answers

Question 499

Explanation

Correct Answer

Incorrect Answers

Question 500

Explanation

Correct Answer

Incorrect Answers

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