Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 43

Question 420

A dynamic inference route picks the cheapest model even when quality fails. Which performance correction follows from the described bottleneck?

Include task quality thresholds and fallback in routing policy.
Always pick the lowest-input-token price regardless of task complexity.
Use generated confidence alone as approval for critical tasks with no calibration.
Remove quality telemetry after routing to reduce operational overhead.

Explanation

Correct Answer

Route by verified task-quality thresholds and escalate/fallback when the cheapest model cannot meet acceptance criteria.

Incorrect Answers

Cheap but incorrect outputs can increase business risk and overall cost.

A model's confidence is not a validated quality threshold.

Without monitored quality, routing regressions go undetected.

Question 421

An enterprise app calls the same external knowledge API for every token. What improves this path without silently weakening answer quality?

Call the same external search API for each generated token in an otherwise static answer.
Cache confidential lookup results globally across unrelated user permissions.
Drop source version identifiers in cached retrieval hits to save space.
Cache source retrieval per request and batch related lookups.

Explanation

Correct Answer

Cache/reuse eligible source retrieval within an authorized request and batch independent lookups rather than calling the external API per output token.

Incorrect Answers

Redundant calls multiply latency, cost and quota use.

Cross-user caching risks information leakage.

The result may become stale or unauditable when sources change.

Question 422

A user requests a preview while heavy processing continues. Which measurement or tuning step should happen first?

Render 'completed successfully' as soon as background processing starts.
Hide all progress for a multi-minute job and assume the user will wait on an open socket.
Stream valid interim status without falsely claiming the transaction is final.
Return guessed final data as a preview without marking it provisional.

Explanation

Correct Answer

Stream an explicit provisional progress/status event while keeping final transaction success separate until authoritative work completes.

Incorrect Answers

Pending work is not a committed outcome.

A durable status contract enables safe continuation and recovery.

Unverified projections can be mistaken for confirmed results.

Question 423

An AI endpoint's p99 latency worsens despite stable model tokens. Which performance correction follows from the described bottleneck?

Scale the model solely because aggregate latency increased without stage-level evidence.
Use token count as a complete explanation of every service p99 spike.
Trace dependency/service network latency and downstream tool spans.
Disable correlation IDs and compare only raw request duration charts.

Explanation

Correct Answer

Trace dependency network, agent tools and backend spans when p99 worsens despite stable model token counts.

Incorrect Answers

The bottleneck may be downstream network/tool waits.

Stable model tokens can coexist with slow external dependencies.

Uncorrelated totals cannot localize the failing dependency.

Question 424

A configuration change raises timeout error rates. What improves this path without silently weakening answer quality?

Profile model API and tool time budgets before raising all timeouts.
Raise every timeout globally to an arbitrary large number after a regression.
Increase only the FM token limit when the failing stage is an external tool.
Report all timeouts as successful transactions to improve alert metrics.

Explanation

Correct Answer

Profile each API/tool's supported deadline, timeout configuration, retries and downstream latency before selectively changing limits.

Incorrect Answers

Longer waits can mask hung services and violate user deadlines.

Generation capacity does not fix tool integration timeouts.

Suppressing errors misrepresents actual business outcomes.

Question 425

A model-cascading optimization routes all hard cases through two models. Which measurement or tuning step should happen first?

Run the smaller model and then always invoke the larger model for every request regardless of the first outcome.
Use the cheapest model's confidence without validating which difficult cases it misroutes.
Measure escalation savings versus extra latency and wrong-branch risk.
Compute the cascade price from only the first model's tokens, ignoring every escalated second invocation.

Explanation

Correct Answer

A two-model cascade should escalate only a measured minority of hard cases; if nearly every input runs both models, added latency/cost can outweigh any savings and risk incorrect routing.

Incorrect Answers

A forced second invocation eliminates the intended escalation savings and adds latency.

Uncalibrated confidence may route challenging cases to an inadequate model.

The extra model call must be included in true cost and tail latency analysis.

Question 426

A GenAI service dashboard shows only HTTP uptime. What observability change gives the team actionable diagnosis?

Declare the AI application healthy whenever the API returns 200 even if answers are wrong.
Add response quality, groundedness, latency and usage metrics with business impact.
Measure only average token-generation rate and ignore user-visible answer completeness.
Count source citations without testing whether passages support the generated claims.

Explanation

Correct Answer

Track grounded response correctness, business outcome, latency and token usage as separate signals; HTTP uptime alone cannot measure GenAI usefulness.

Incorrect Answers

Successful transport does not imply factual correctness or task success.

Throughput is not grounded answer quality.

Citation presence does not establish factual entailment.

Question 427

A team cannot attribute high token cost to application owner. Which monitoring signal or correlation is missing?

Attribute every Bedrock request to one common billing account with no request tags or caller identity.
Infer application owner from the model's natural-language answer.
Use invocation request metadata or principal identity and aggregate token usage.
Count only the number of calls and assume equal token cost across all owners.

Explanation

Correct Answer

Use per-invocation requestMetadata or captured IAM principal identity, then aggregate measured input/output tokens by accountable application owner.

Incorrect Answers

A common charge bucket cannot identify which application caused spend.

Model output does not carry trusted billing attribution.

Request sizes and model prices vary, so call counts misallocate expense.

Question 428

A model's output token count doubles after a prompt change. Which telemetry control best identifies the failure?

Average tokens across all historical versions and ignore the released prompt ID.
Use only overall account billing totals without separating input and output.
Remove token telemetry to lower logging cost during a usage spike.
Compare version-tagged token distributions and alert on anomaly.

Explanation

Correct Answer

Compare exact prompt-version-tagged input/output usage distributions and alert on sustained output-token anomalies after deployment.

Incorrect Answers

Version mixing hides which change increased generation length.

The diagnostic requires identifying output-token growth.

Without usage counters, the team cannot measure or verify the regression.

Question 429

An application sees bursts of invalid guardrail responses. What observability change gives the team actionable diagnosis?

Treat every rise in blocked responses as proof guardrails improved without checking allowed requests.
Ignore the precise guardrail version used for each blocked inference.
Disable guardrails whenever the block count spikes.
Track block reason categories, source/prompt changes and false-block impact.

Explanation

Correct Answer

Segment guardrail interventions by policy/version/block reason and evaluate false blocks on legitimate cases to distinguish safety incidents from overblocking.

Incorrect Answers

A bad policy update could be rejecting benign customers.

Version drift prevents diagnosis of category changes.

Removing safety controls is not a controlled triage or remediation.

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.

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

Question 420

Explanation

Correct Answer

Incorrect Answers

Question 421

Explanation

Correct Answer

Incorrect Answers

Question 422

Explanation

Correct Answer

Incorrect Answers

Question 423

Explanation

Correct Answer

Incorrect Answers

Question 424

Explanation

Correct Answer

Incorrect Answers

Question 425

Explanation

Correct Answer

Incorrect Answers

Question 426

Explanation

Correct Answer

Incorrect Answers

Question 427

Explanation

Correct Answer

Incorrect Answers

Question 428

Explanation

Correct Answer

Incorrect Answers

Question 429

Explanation

Correct Answer

Incorrect Answers

Question 430

Explanation

Correct Answer

Incorrect Answers

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