Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 14

Question 130

A high-recall query returns unrelated departments. Which correction most directly fixes this grounding failure?

Raise the number of retrieved chunks without restricting the department.
Prompt the FM to ignore the wrong department after the restricted chunks entered its context.
Change the text embedding model without defining allowed business-domain metadata.
Constrain search by tenant/business domain and evaluate precision at k.

Explanation

Correct Answer

Apply domain and permission metadata filters and measure precision at k; broad dense similarity alone can cross department boundaries.

Incorrect Answers

More results can increase unrelated department content.

Filtering after exposure does not enforce source authorization and cannot correct ranking inputs.

A new embedding representation does not implement domain scope or rights.

Question 131

Retrieval latency rises sharply from large semantic candidates. What retrieval mechanism or check should be selected for this case?

Tune top-k, ANN/index parameters and rerank only bounded candidates.
Rerank thousands of candidates on every request without measuring latency.
Reduce candidates to one without evaluating whether the correct document remains in recall.
Increase response token limits to compensate for slow retrieval queries.

Explanation

Correct Answer

Reduce or tune ANN candidate top-k and rerank only a bounded shortlist while validating recall and p95 latency against gold queries.

Incorrect Answers

Unbounded reranking amplifies the bottleneck this case needs to fix.

Aggressive pruning may improve speed while destroying the task's evidence coverage.

Generation length does not address candidate-search latency.

Question 132

A document has figures requiring spatial interpretation. What should be done before treating retrieved content as authoritative evidence?

Index only OCR-extracted plain text and discard all figure positions and labels.
Ask the model to describe figures using document titles without the figure content.
Return the entire image without verifying the selected model's multimodal input/size support.
Use supported multimodal parsing or associated caption content for grounding.

Explanation

Correct Answer

Use a supported multimodal parser or associated figure/caption context with provenance so the model can reason about diagram layout.

Incorrect Answers

The spatial relationships in the source cannot be reliably recovered from plain text alone.

A title is insufficient evidence of arrows, placement and numeric labels.

Unsupported formats or sizes fail rather than guaranteeing grounded interpretation.

Question 133

A user needs exact numeric totals from tables. Which correction most directly fixes this grounding failure?

Ask the model to estimate totals from nearby row descriptions instead of numeric values.
Sum embedded vector similarities and treat them as spreadsheet amounts.
Retrieve row/cell evidence then compute totals deterministically.
Omit units/currency from the retrieved cells and present one aggregated number.

Explanation

Correct Answer

Retrieve the actual numeric cells and units, then run deterministic arithmetic; model prose is not a trusted calculator.

Incorrect Answers

Narrative similarity does not determine exact numerical totals.

Similarity scores are not table cell values or a valid accounting operation.

Mixing units without conversion and provenance makes the result unverifiable.

Question 134

A KB continuously ingests source updates but client cache remains stale. What retrieval mechanism or check should be selected for this case?

Increase Knowledge Base sync frequency but never expire the client-side cached answers.
Rely on the model's training cutoff to distinguish newly updated source documents.
Align index synchronization with cache invalidation and freshness tests.
Clear only the server logs while leaving the retrieval response cache intact.

Explanation

Correct Answer

Source synchronization and application cache freshness are separate dependencies; invalidate cached answers whenever underlying indexed source versions change.

Incorrect Answers

A fresh index will not help if every request is served from an older application cache.

Pretraining contains no live cache invalidation signal.

Log rotation does not replace stale cached evidence.

Question 135

An orchestration engine sends full document hits repeatedly. What should be done before treating retrieved content as authoritative evidence?

Pass full duplicated source text to every orchestration stage without caching safe identifiers.
Normalize a retrieval result schema and cache eligible source identifiers.
Cache responses globally without caller/document permission or version keys.
Drop document IDs and retain only the generation text to reduce bytes.

Explanation

Correct Answer

Normalize retrieval hits into a bounded contract with source/version IDs; reuse permitted references when identical evidence is requested within valid freshness scope.

Incorrect Answers

Repeated long passages waste context and make source/version consistency harder to check.

An unscoped cache may leak restricted documents or return stale revisions.

Removing IDs destroys auditability of which evidence the workflow used.

Question 136

A shared prompt has hardcoded customer names. Which action preserves the correct instruction or output boundary?

Edit a single shared prompt draft for each customer and publish it immediately.
Concatenate customer input into the system role without escaping or validation.
Replace hardcoded text with typed prompt variables and validate allowed substitution.
Hardcode every customer's name in a separate model deployment and skip variable validation.

Explanation

Correct Answer

Typed prompt variables keep customer values separate from stable instructions; validate substitutions for required type, scope and permitted encoding.

Incorrect Answers

Mutating the shared template mixes tenants and makes releases irreproducible.

Untrusted values can break template structure or masquerade as instructions.

Per-customer hardcoding causes drift and does not establish safe substitution rules.

Question 137

A team edits production prompt text without version control. Which change is necessary before this prompt flow is promoted?

Let production call the mutable draft prompt ID with no pinned release version.
Log only the editor's initials while overwriting the previous prompt content.
Publish immutable Prompt Management versions and approve promotion.
Version just the UI copy while leaving Bedrock prompt variables and model config mutable.

Explanation

Correct Answer

Bedrock Prompt Management version snapshots fix tested prompt text and configuration for controlled promotion and rollback.

Incorrect Answers

A draft change may silently alter live output without promotion evidence.

An editor name cannot reconstruct the tested prompt or its inference parameters.

The effective runtime prompt may differ from the reviewed text.

Question 138

Two prompt variants produce different costs and accuracy. Which prompt or application control best resolves the requirement?

Compare the cheaper prompt on easier questions and the longer prompt on harder questions.
Choose the variant solely by total output token count without checking task errors.
Use a single fluent example response as proof of aggregate quality.
Compare variants on the same held-out dataset and token budget.

Explanation

Correct Answer

Evaluate prompt variants on identical held-out cases with comparable token budgets, correctness and latency before picking the cheaper one.

Incorrect Answers

Different evaluation populations cannot establish a fair accuracy/cost tradeoff.

Low token usage is not an acceptable substitute for correct task performance.

One anecdote does not measure reliability across representative cases.

Question 139

A system instruction is mistakenly included as ordinary user content. Which action preserves the correct instruction or output boundary?

Prefix a user message with the word SYSTEM and assume it changes role authority.
Use the appropriate message roles to preserve instruction hierarchy.
Place retrieved documents before trusted instructions in the system block.
Rely on a post-response regex to repair every instruction-hierarchy mistake.

Explanation

Correct Answer

Place trusted system-level instructions in supported message roles and user data in its proper role, rather than expecting text labels to enforce hierarchy.

Incorrect Answers

Literal text labels inside user content do not become an API system role.

Untrusted source data should not be elevated above trusted rules.

Regex output cleanup cannot reliably prevent model actions driven by mis-scoped instructions.

Question 140

A model generates unrecognized status codes. Which change is necessary before this prompt flow is promoted?

Accept any syntactically valid JSON string as an approved business status.
Require a structured response contract and validate the enum downstream.
Increase generation temperature until one output happens to match an allowed status.
Treat an unspecified fallback string as equivalent to an approved state in downstream code.

Explanation

Correct Answer

Constrain output to the supported schema where available and validate enum values in application code; syntax correctness alone does not ensure business validity.

Incorrect Answers

A valid string may still be an unauthorized enum value.

Random sampling does not enforce a stable contract.

Unknown states must be rejected or handled explicitly, not silently authorized.

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

Question 130

Explanation

Correct Answer

Incorrect Answers

Question 131

Explanation

Correct Answer

Incorrect Answers

Question 132

Explanation

Correct Answer

Incorrect Answers

Question 133

Explanation

Correct Answer

Incorrect Answers

Question 134

Explanation

Correct Answer

Incorrect Answers

Question 135

Explanation

Correct Answer

Incorrect Answers

Question 136

Explanation

Correct Answer

Incorrect Answers

Question 137

Explanation

Correct Answer

Incorrect Answers

Question 138

Explanation

Correct Answer

Incorrect Answers

Question 139

Explanation

Correct Answer

Incorrect Answers

Question 140

Explanation

Correct Answer

Incorrect Answers

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