Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 44

Question 430

A Bedrock API is returning ThrottlingException but no alarm exists. Which monitoring signal or correlation is missing?

Treat every 429 as a model hallucination rather than a capacity/quotas problem.
Increase output token budgets while failing requests are already rate-limited.
Only alarm on EC2 CPU because Bedrock is accessed through an API.
Monitor invocation errors and quota utilization with CloudWatch alerts.

Explanation

Correct Answer

Monitor Bedrock ThrottlingException and measured quota/token usage with actionable CloudWatch alarms and bounded retry/traffic controls.

Incorrect Answers

A throttling error happens before a successful generated answer.

Larger demands can worsen quota pressure.

Managed model invocation quotas/errors need their own signals.

Question 431

A tool orchestration failure is misreported as FM hallucination. Which telemetry control best identifies the failure?

Label a failed Lambda invocation as hallucination because the FM mentioned a tool.
Measure only the final response length without collecting tool duration and status.
Trace agent-to-tool-to-service spans to locate actual failing component.
Discard intermediate tool errors and trust the supervisor agent's final summary.

Explanation

Correct Answer

Use correlated spans for agent plan, tool invocation and backend transaction to establish whether the failure was in orchestration, tool or generation.

Incorrect Answers

A backend execution error is not necessarily model text fabrication.

Length cannot identify the failing service call.

A summary may misstate an actual failed or partial side effect.

Question 432

An application successfully invokes a supported Bedrock runtime model in one AWS account and Region, but the team has never enabled model invocation logging for that account and Region. What observability setup must be completed before expecting invocation records?

Enable model invocation logging for the invoking account and Region and configure valid CloudWatch Logs and/or S3 destinations for supported runtime calls.
Assume every bedrock-mantle call appears in bedrock-runtime invocation logs.
Check only the existence of a CloudWatch log group without verifying the region's logging configuration.
Grant a foundation model more tokens as the remedy for missing telemetry.

Explanation

Correct Answer

Model invocation logging is not enabled by default. The configured destinations and access permissions apply within the invoking account/Region and supported runtime operations; creating an unrelated application log group does not enable this service feature.

Incorrect Answers

AWS documents endpoint-specific logging support.

An empty group cannot prove invocation logging was enabled for those operations.

Generation settings do not configure logging capture.

Question 433

An application uses a different account's S3 log bucket. Which monitoring signal or correlation is missing?

Send Bedrock invocation logs directly to an arbitrary bucket in another account/Region as though it were always supported.
Respect same-account and Region logging destination constraints where required.
Use a public bucket policy so the logger can write anywhere without account checks.
Rename the S3 bucket while leaving the destination's physical Region unchanged.

Explanation

Correct Answer

Invocation logging delivery destinations must satisfy the account/Region constraints and permissions of the actual logging configuration.

Incorrect Answers

The service's direct destination requirements constrain this setup.

Broad public access does not implement a secure logging destination.

A name change cannot resolve a regional mismatch.

Question 434

A logging pipeline stores sensitive prompt text by default. Which telemetry control best identifies the failure?

Minimize/sanitize payloads and enforce destination IAM, encryption and retention.
Write every raw customer prompt to a broadly accessible dashboard for debugging.
Base64-encode sensitive log data and classify it as sanitized.
Set a short UI log-view TTL while S3 keeps the raw logs indefinitely.

Explanation

Correct Answer

Minimize sensitive prompt bodies, tightly scope log destination IAM and apply encryption/retention; logs can contain full model inputs/outputs.

Incorrect Answers

Invocation logging can leak personal data and proprietary text.

Encoding is reversible and not data protection.

Front-end expiry does not enforce backend object deletion.

Question 435

A Bedrock invocation log contains oversized multimodal input. What observability change gives the team actionable diagnosis?

Check whether large/binary data was delivered via S3 rather than inline CloudWatch event.
Assume every image byte and JSON body over 100 KB is fully included in one CloudWatch event.
Treat a small log event as proof the model never received the image.
Copy sensitive large document bodies into plain-text CloudWatch messages to avoid S3 references.

Explanation

Correct Answer

For Bedrock runtime logs, binary/large inputs can be stored as separate S3 objects; inspect the recorded S3 reference rather than expecting all bytes inline in CloudWatch.

Incorrect Answers

AWS stores large/binary invocation payloads separately in S3 when configured.

The event can point to larger S3 payload storage.

This defeats supported payload handling and can leak confidential inputs.

Question 436

A model route changes but dashboard merges all model IDs. Which monitoring signal or correlation is missing?

Aggregate all FM responses under one 'AI' metric series regardless of active model/version.
Segment errors and quality by exact model ID/inference profile/version.
Change route labels but keep the same unsplit error histogram.
Assume every response from a model provider has identical semantics across releases.

Explanation

Correct Answer

Segment errors, latency, usage and content quality by exact model ID, deployment profile and version to identify regression after routing changes.

Incorrect Answers

Different routes and versions can hide model-specific failures.

Cosmetic renaming cannot isolate the degrading model.

Versions can differ in output contract, modality and quality.

Question 437

A vector store silently loses index freshness. Which telemetry control best identifies the failure?

Monitor ingestion lag, source/index count and stale retrieval tests.
Declare retrieval healthy solely because the vector endpoint returns HTTP 200.
Run one latency-only query and ignore document revision metadata.
Delete cached results only when a customer complains of stale policy answers.

Explanation

Correct Answer

Monitor source/index synchronization lag, indexed document counts and a versioned stale-hit test; vector-store uptime alone misses freshness failures.

Incorrect Answers

A live endpoint can return outdated chunks.

Speed does not establish that new source content is indexed.

Reactive cache clearing does not monitor ingestion correctness.

Question 438

An agent tool has unpredictable p95 duration. What observability change gives the team actionable diagnosis?

Track total chat duration only and assign every delay to model inference.
Use mean tool duration while ignoring rare severe timeouts.
Increase model temperature to improve the tool's server execution time.
Track tool call counts, timeouts and per-tool duration distributions.

Explanation

Correct Answer

Measure per-tool call count, p95/p99 duration, timeout, retry and downstream status, separating tool latency from model decode.

Incorrect Answers

An external tool may dominate the tail independently of FM latency.

The p95 problem may be hidden by average statistics.

Sampling does not change backend tool response latency.

Question 439

A multi-agent workflow frequently repeats tasks. Which monitoring signal or correlation is missing?

Allow specialists to run the same non-idempotent write concurrently to finish faster.
Count final agent messages without recording repeated intermediate tool invocations.
Analyze coordination traces and deduplicate duplicate tool invocations.
Turn off collaboration traces after a supervisor returns success.

Explanation

Correct Answer

Correlate multi-agent execution traces and stable task IDs to detect duplicate planning and tool calls; dedupe safe work before repeated effects.

Incorrect Answers

Duplicate agents can cause repeated charges or state conflicts.

Output counts cannot detect wasteful duplicate tool execution.

The observed response does not prove coordination was efficient.

Question 440

A response shows citations but support rate is falling. Which telemetry control best identifies the failure?

Count hyperlinks in generated answers and call the proportion grounded accuracy.
Run grounded-answer quality evaluation on a stable verified question set.
Use only retrieval similarity scores to assess whether generation faithfully used evidence.
Change the benchmark questions every release and compare raw aggregate support rates.

Explanation

Correct Answer

Evaluate whether each cited passage entails the corresponding answer claim on a stable verified reference set; citation presence alone is insufficient.

Incorrect Answers

A valid URL can support a different claim.

Search ranking does not test statement entailment.

Uncontrolled test population changes confound the trend.

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

Question 430

Explanation

Correct Answer

Incorrect Answers

Question 431

Explanation

Correct Answer

Incorrect Answers

Question 432

Explanation

Correct Answer

Incorrect Answers

Question 433

Explanation

Correct Answer

Incorrect Answers

Question 434

Explanation

Correct Answer

Incorrect Answers

Question 435

Explanation

Correct Answer

Incorrect Answers

Question 436

Explanation

Correct Answer

Incorrect Answers

Question 437

Explanation

Correct Answer

Incorrect Answers

Question 438

Explanation

Correct Answer

Incorrect Answers

Question 439

Explanation

Correct Answer

Incorrect Answers

Question 440

Explanation

Correct Answer

Incorrect Answers

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