Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 42

Question 410

An FM gets slower as conversation history grows. Which measurement or tuning step should happen first?

Send the entire multi-year conversation to each FM call indefinitely.
Prune/summarize context using supported retained state without losing key facts.
Delete all earlier session facts without keeping necessary case state.
Use one globally shared summary for unrelated actors' conversations.

Explanation

Correct Answer

Prune irrelevant history and persist authorized summaries/critical state outside the prompt, measuring quality while reducing context length.

Incorrect Answers

Long histories increase tokens and eventually exceed the context window.

Missing essential details can create inconsistent answers or tool mistakes.

Cross-user state pooling is a privacy and correctness defect.

Question 411

A model performs classification at high temperature. Which performance correction follows from the described bottleneck?

Increase temperature to maximize randomness on a fixed-label classification task.
Assume temperature zero guarantees identical outputs across all model versions.
Approve the classifier solely because labels have valid JSON syntax.
Tune deterministic-friendly inference parameters and verify actual accuracy.

Explanation

Correct Answer

Use task-appropriate temperature/sampling settings and evaluate classification accuracy; low temperature improves repeatability but cannot guarantee correctness.

Incorrect Answers

Sampling variation can reduce consistency without adding evidence.

Backend/model differences can still change results.

Format does not establish correct class assignment.

Question 412

Two candidate model versions produce comparable validated answer quality but different p95 response times under load. Which TWO steps support a fair release decision against the production latency SLO?

Measure the first token on one idle request for each model and treat it as full response latency.
Benchmark both versions with the same representative prompt mix and realistic concurrent request load.
Select the cheaper version without evaluating its user-visible response time or quality acceptance threshold.
Compare observed p95 end-to-end completion against the specified latency and quality release thresholds.
Make the faster model appear to pass by dropping its slowest requests from the evaluation dataset.

Explanation

Correct Answer

Comparable concurrency and input/output demands reveal queuing and model-speed differences without changing the workload.

A valid decision must satisfy both the measured tail-latency SLO and the accepted-answer quality gate.

Incorrect Answers

Time to first token omits output generation, retrieval and other processing needed to deliver a usable answer.

Price alone cannot establish adherence to documented latency and quality objectives.

Excluding slow or timed-out requests biases p95 and conceals real user experience.

Question 413

A workload is optimized for latency but has unstable final output. Which measurement or tuning step should happen first?

Select the fastest generation even when repeated answers contradict each other.
Treat every different word choice as a semantic contradiction.
Evaluate repeatability and quality, not only throughput.
Disable output validation to avoid extra milliseconds while accepting incorrect results.

Explanation

Correct Answer

Benchmark repeated predictions and factual quality under the same inputs/models, not only token throughput; latency improvements may hide unstable answers.

Incorrect Answers

Fast responses do not satisfy stable correctness requirements.

Harmless stylistic variation should not count as inconsistency.

Faster unsafe answers are not an acceptable optimization.

Question 414

A vector index returns results slowly after rapid growth. Which performance correction follows from the described bottleneck?

Tune ANN/search index and shard organization with monitored recall and QPS.
Double index shards blindly without measuring query pattern or recall.
Drop search accuracy validation and select the absolute smallest top-k.
Increase the answer model's output tokens while leaving slow vector search untouched.

Explanation

Correct Answer

Tune ANN parameters and shard/index layout against measured recall and throughput, distinguishing scale-related latency from missing evidence.

Incorrect Answers

Extra shards can increase coordination overhead and latency.

Aggressive pruning may miss relevant documents.

Generation length does not address index search latency.

Question 415

A hybrid search query returns too many irrelevant lexical hits. What improves this path without silently weakening answer quality?

Raise every lexical field boost to maximum without testing whether domain results worsen.
Disable all vector similarity and trust exact word overlap for paraphrased questions.
Use a global prompt rewrite to hide irrelevant retrieved passages after they enter context.
Adjust field boosts and ranking; measure relevance and retrieval latency.

Explanation

Correct Answer

Tune hybrid lexical/vector weighting and field boosts on labeled queries, ensuring filterable text and store support remain valid.

Incorrect Answers

Excess keyword weight can overwhelm semantically relevant results.

Pure lexical matching can hurt recall for synonymous requests.

Post-hoc prompt cleanup cannot fully correct candidate ranking.

Question 416

An inference app sees queueing spikes at peak hours. Which measurement or tuning step should happen first?

Plan token throughput and concurrent in-flight request limits from measured load.
Count only requests per minute while ignoring that some replies generate ten times more tokens.
Allow unlimited simultaneous requests because AWS will scale every model instantly.
Increase frontend timeout and ignore a growing backend queue.

Explanation

Correct Answer

Capacity plan token throughput, concurrency, queue admission and model response length under realistic peak load to avoid tail-queue spikes.

Incorrect Answers

Token demand and active generation time determine queue saturation.

Service quotas and provisioning delay can cause severe throttling.

Longer waits do not reduce contention or served tokens per second.

Question 417

A long output generation occupies serving capacity. Which performance correction follows from the described bottleneck?

Allow long unbounded essays for one-word classification while other callers queue.
Apply output size limits or route simple tasks to shorter response patterns.
Truncate responses mid-JSON without checking required fields.
Remove safety, grounding and business context just to save output tokens.

Explanation

Correct Answer

Bound response lengths and route simple tasks to concise validated formats, monitoring whether completion quality remains acceptable.

Incorrect Answers

Excess generation occupies model throughput.

Aggressive output caps can produce unusable incomplete data.

Reducing critical instructions can degrade correctness.

Question 418

A model cannot serve a sudden burst of requests. What improves this path without silently weakening answer quality?

Issue unlimited retries immediately during a demand spike.
Assume one static instance handles arbitrarily many concurrent token generations.
Implement bounded concurrency, autoscaling or queue admission controls.
Accept requests and report them completed before any model execution.

Explanation

Correct Answer

Enforce admission control and bounded queues with supported autoscaling/capacity while returning a truthful backpressure status when saturated.

Incorrect Answers

Immediate unbounded retries compound the overload by creating new requests while the service is throttled, rather than reducing admitted concurrency.

Physical compute and service quota limits remain finite.

Queue acceptance is not generated answer success.

Question 419

A GPU endpoint is memory-bound despite low CPU. Which measurement or tuning step should happen first?

Add CPU cores without examining GPU out-of-memory or KV-cache pressure.
Monitor model memory/throughput and choose supported hosting configuration.
Increase batch/concurrency indefinitely while VRAM is already exhausted.
Select a new instance type solely from its hourly price without checking model fit.

Explanation

Correct Answer

Profile GPU VRAM, memory bandwidth, token throughput and supported serving settings; CPU idleness does not establish GPU headroom.

Incorrect Answers

CPU scaling may not address the actual accelerator bottleneck.

More in-flight work can worsen memory failures.

Model-serving compatibility and memory limits are mandatory.

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.

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

Question 410

Explanation

Correct Answer

Incorrect Answers

Question 411

Explanation

Correct Answer

Incorrect Answers

Question 412

Explanation

Correct Answer

Incorrect Answers

Question 413

Explanation

Correct Answer

Incorrect Answers

Question 414

Explanation

Correct Answer

Incorrect Answers

Question 415

Explanation

Correct Answer

Incorrect Answers

Question 416

Explanation

Correct Answer

Incorrect Answers

Question 417

Explanation

Correct Answer

Incorrect Answers

Question 418

Explanation

Correct Answer

Incorrect Answers

Question 419

Explanation

Correct Answer

Incorrect Answers

Question 420

Explanation

Correct Answer

Incorrect Answers

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