Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 13

Question 120

A KB citation links to an inaccessible internal bucket. What should be done before treating retrieved content as authoritative evidence?

Expose s3://private-bucket/key in the answer and claim it is clickable evidence.
Make the entire bucket public so the assistant's citations always work.
Provide authorized citation links, not raw private objects to end users.
Replace the source link with a fake public URL without checking what it serves.

Explanation

Correct Answer

A citation must use an authorized reader path, such as a checked application route, not reveal an inaccessible private S3 object URI.

Incorrect Answers

A raw private object locator is neither a user-authorized URL nor a readable citation.

This leaks unrelated private documents and defeats object-level access controls.

Invented links do not provide traceable evidence and may misrepresent content access.

Question 121

A long reference document overwhelms the answer context. Which correction most directly fixes this grounding failure?

Pass all document chunks into a prompt that exceeds the model context window.
Limit retrieved result count and token budget while retaining sufficient evidence.
Drop every source snippet and rely on the model's pretraining for policy answers.
Truncate the final generated output while leaving a multi-thousand-token irrelevant retrieval payload.

Explanation

Correct Answer

Bound retrieved source count and token footprint while preserving the passages that answer the question; verify supported evidence coverage.

Incorrect Answers

Unbounded context risks rejection or truncation and reduces dependable source coverage.

A short prompt without cited current evidence cannot satisfy grounded-answer requirements.

Output limits do not control an oversized input context.

Question 122

An indexing system changes embedding model generation. What retrieval mechanism or check should be selected for this case?

Mix old corpus embeddings with new query vectors because both arrays have the same length.
Change only the k-NN top-k setting while leaving vector model generations inconsistent.
Re-rank incompatible nearest-neighbor results without measuring missed gold documents.
Re-embed stored documents and evaluate old-vs-new retrieval before cutover.

Explanation

Correct Answer

A model-generation switch changes embedding space; reindex the corpus with the new model or use the old model consistently for both queries and vectors.

Incorrect Answers

Equal vector dimensionality does not guarantee comparable semantic spaces.

Retrieval tuning does not repair incompatible representations.

Reranking cannot recover relevant documents excluded from the candidate set by invalid embeddings.

Question 123

A Spanish-language query is being mistranslated before search, causing relevant policy documents to be missed. Select TWO comparative evaluations of feasible multilingual retrieval paths before choosing a production route.

Automatically translate every query to English without measuring changes in entity/meaning.
Test multilingual direct-embedding quality.
Change only the vector search top-k regardless of language mismatch.
Declare one route better after testing different documents and questions for each.
Compare with an approved translate-before-search retrieval path.

Explanation

Correct Answer

Test direct multilingual retrieval on labeled language pairs, including domain idioms, before committing to a model path.

Compare a separately validated translation+retrieval path against the same gold corpus without assuming automatic translation accuracy.

Incorrect Answers

Translation can alter technical terms and query intent; untested translation is not a verified remedy.

More candidates do not prove a cross-language embedding model covers the requested locale.

A confounded comparison cannot isolate which retrieval path actually performs better.

Question 124

A Bedrock Knowledge Base uses hierarchical chunking. A retrieval request targets eight child matches, but those matches collapse to three distinct parent passages. Source-object counts and ingestion health remain unchanged. How should the architect interpret the smaller final result count?

Run a full re-ingestion immediately just because fewer result objects were returned.
Account for child-to-parent replacement and deduplication, rather than treating fewer final parent passages as evidence of lost source documents.
Assume numberOfResults equals distinct final parent chunks exactly.
Delete parent-child linkage metadata to force matching result counts.

Explanation

Correct Answer

In hierarchical chunking, several retrieved child hits can map to the same parent chunk, which is returned once. The configured retrieval candidate count does not guarantee an equal number of distinct final parent passages.

Incorrect Answers

A normal hierarchical parent-collapse can explain the count without missing documents.

This setting constrains child retrieval; shared parent chunks can reduce final returned item count.

Removing hierarchy sacrifices the parent context that the application deliberately configured.

Question 125

A no-chunk KB configuration loses page-number citation metadata. What retrieval mechanism or check should be selected for this case?

Choose supported chunking strategy when page citations are required.
Keep no-chunk mode and fabricate page numbers from a model-generated guess.
Increase Retrieve numberOfResults without changing the no-chunk ingestion strategy.
Add a display label reading page 1 to every result regardless of actual document pagination.

Explanation

Correct Answer

No-chunk mode treats the document as one chunk and does not provide the page-number metadata required for page-specific citations.

Incorrect Answers

Source page citations require indexed page metadata, not a language-model estimate.

More results cannot create missing page-number metadata.

A constant label misrepresents the source page location.

Question 126

A policy RAG assistant retrieves clauses from several revisions and cites the wrong one. Select THREE source-version handling steps that preserve provenance, select the effective revision, and display that revision in the answer.

Preserve authoritative document revision metadata.
Filter retrieval to effective policy revision.
Sort all retrieved clauses by vector similarity and disregard their effective dates.
Hide version fields in the UI while leaving superseded source chunks searchable.
Carry revision and date into cited answer.
Ask the model to infer current policies from names of documents without source metadata.

Explanation

Correct Answer

Store and filter source effective dates to select valid revisions before generation.

Preserve revision/version provenance so a retrieved clause can be mapped to the actual approved source.

Reject expired or superseded clauses using evidence of their validity, not semantic ranking alone.

Incorrect Answers

Semantic relevance does not establish which policy revision is legally applicable.

A cosmetic display change does not prevent the model from grounding on revoked revisions.

Model inference cannot replace authoritative effective-date and approval checks.

Question 127

An assistant retrieves an internal policy passage through Bedrock Knowledge Bases, but that passage contains a forged SYSTEM instruction to disclose other customers’ records. The application currently interpolates the entire passage as if it were privileged prompt instructions. What trust-boundary correction prevents retrieval-based instruction takeover?

Place retrieved paragraphs in the system-message role so any embedded command is obeyed.
Execute tool directives contained in retrieved pages before verifying the user and operation.
Treat retrieved Knowledge Bases passages as untrusted evidence, never as system/developer instructions, and prohibit any tool-policy override originating from that passage.
Disable prompt-injection testing once the document comes from an indexed corporate source.

Explanation

Correct Answer

Knowledge Base documents supply evidence, not authority. Keep retrieved content in a lower-trust context and validate any requested actions against independent identity/tool permissions.

Incorrect Answers

This elevates potentially attacker-controlled text into the trusted instruction channel.

Source content cannot grant capability or actor authorization for external actions.

Compromised or malicious source text can still contain hostile instructions after ingestion.

Question 128

A retrieval store returns graph nodes but the prompt loses relationships. What retrieval mechanism or check should be selected for this case?

Give the model a flat list of graph node names with no edge direction or relationship type.
Rank graph nodes by embedding distance and assume rank means causation.
Delete edge provenance and let the model invent missing paths between entities.
Preserve entity/edge context in grounded evidence assembly.

Explanation

Correct Answer

Graph-grounded answers require the relevant entity-to-entity edges and provenance, not just isolated vector hits for node labels.

Incorrect Answers

Entity names alone cannot establish the required relationships.

Similarity does not encode directed graph semantics.

A generated path without supported edges is not an evidence-backed relationship.

Question 129

A search tool times out during interactive response. What should be done before treating retrieved content as authoritative evidence?

Respond from stale cached snippets without checking freshness or authorization after timeout.
Retry the failing endpoint indefinitely before returning any status.
Treat a timed-out retrieval as an empty successful result and invent a source citation.
Use bounded retry/circuit-breaker and truthful fallback rather than stale invented evidence.

Explanation

Correct Answer

Use bounded retry and circuit breaking, then return a truthful unavailable/pending answer if search evidence cannot be obtained.

Incorrect Answers

Cached data may be obsolete or out of scope and cannot be claimed as a fresh retrieval.

Unbounded retries violate interactive deadlines and can amplify service failure.

A timeout means unknown retrieval state, not proof that no supporting document exists.

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.

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

Question 120

Explanation

Correct Answer

Incorrect Answers

Question 121

Explanation

Correct Answer

Incorrect Answers

Question 122

Explanation

Correct Answer

Incorrect Answers

Question 123

Explanation

Correct Answer

Incorrect Answers

Question 124

Explanation

Correct Answer

Incorrect Answers

Question 125

Explanation

Correct Answer

Incorrect Answers

Question 126

Explanation

Correct Answer

Incorrect Answers

Question 127

Explanation

Correct Answer

Incorrect Answers

Question 128

Explanation

Correct Answer

Incorrect Answers

Question 129

Explanation

Correct Answer

Incorrect Answers

Question 130

Explanation

Correct Answer

Incorrect Answers

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