Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 10

Question 90

A Knowledge Base source has duplicated old and new versions of a contract under different object paths. Search returns both as if each were current. How should indexing represent them?

Delete all version metadata and keep whichever chunk has the highest vector score.
Increase the number of duplicate chunks to improve search confidence.
Rename every file to the current date regardless of its actual document effective date.
Deduplicate by stable logical document identity while retaining version lineage and marking which version is authorized for current retrieval.

Explanation

Correct Answer

Document identity and version are different concepts. Version-aware deduplication prevents redundant outdated hits without erasing the audit path to prior contract revisions.

Incorrect Answers

High relevance cannot determine which version is current or preserve traceability.

Indexing the same outdated content more often can crowd out current evidence.

Cosmetic naming does not prove version identity or legal currency.

Question 91

A compliance search index holds confidential source text and vector metadata. Users need access to specific projects, not the entire corpus. Which storage protection is appropriate?

Encrypt the index but grant every employee unrestricted document search access.
Use a common global service identity for all projects without query-time authorization.
Encrypt source and index as required, restrict principals to necessary operations, and apply project/document access rules to searchable metadata.
Remove source metadata while storing private text unchanged and assume privacy improves.

Explanation

Correct Answer

Encryption, principal permissions and retrieval ACLs cover different boundaries; project-sensitive text can leak if any authenticated caller can query all indexed chunks.

Incorrect Answers

At-rest encryption is bypassed legitimately by a broadly authorized querying service; it cannot replace read permissions.

A shared high-privilege role can fetch restricted chunks unless document-level access is enforced.

Deleting provenance does not stop unauthorized access and weakens document-level filters.

Question 92

A document connector ingests HR policies but drops the per-page ACL attributes from the original system. Which preprocessing change keeps the resulting KB permission-aware?

Preserve effective page/document ACL attributes as indexable metadata and enforce them against the requesting user before publishing retrieved chunks.
Permit all authenticated employees to query every page because the source connector itself has read permission.
Remove ACL labels and rely on a prompt telling the FM to ignore HR-only pages.
Use only TLS encryption and skip identity-scoped query filtering.

Explanation

Correct Answer

A connector can lose source permissions unless ACL semantics are mapped explicitly. Metadata and caller authorization must be enforced before results become model context.

Incorrect Answers

A broad ingestion identity does not grant downstream users the original page-level access rights.

The content can already be retrieved, and a prompt is not equivalent to record authorization.

TLS protects in-transit bytes, not which employee may retrieve sensitive indexed content.

Question 93

A Knowledge Base reads source PDFs from S3 but writes embeddings to a separate OpenSearch vector engine. Operations staff assume one S3 policy covers all writes. What permission model is required?

Separately authorize S3 object reads, the Bedrock KB service role and the vector-store index/write data-plane actions.
Grant the S3 bucket owner full permission and assume that automatically grants OpenSearch index writes.
Create a valid embedding model configuration and ignore the storage principal.
Change the source filename extension to JSON to bypass vector engine authorization.

Explanation

Correct Answer

Source storage, Bedrock orchestration and the vector engine are distinct resources. The exact execution identity must hold each necessary permission and satisfy collection network policy.

Incorrect Answers

An S3 bucket policy cannot confer OpenSearch Serverless collection/index data access.

An embedding model can generate vectors that still fail when a service role cannot write them.

File format cannot grant permissions across AWS services or indexes.

Question 94

A vector retriever returns highly similar insurance documents from the wrong product line. The query names a specific policy product. Which retrieval constraint should be added?

Increase the vector similarity threshold without constraining the policy product.
Filter by validated business-domain/product metadata before ranking or generating an answer, then assess precision for that product.
Remove all product metadata from documents before indexing to avoid bias.
Ask the FM to prefer the correct product after including both product lines in context.

Explanation

Correct Answer

Semantic closeness alone cannot enforce which policy corpus is authoritative for a product. Metadata narrowing makes the result set meet the stated business context.

Incorrect Answers

An unrelated but semantically near-identical policy may still score above the threshold.

Deleting the domain field removes the very control needed to exclude wrong-policy results.

The model cannot reliably undo retrieval-domain contamination once conflicting source text is supplied.

Question 95

A migration embedded only half of an OpenSearch vector index with a new embedding model; remaining partitions still contain vectors from the old embedding space. Shard sizes and topology are already healthy. What is the direct correction?

Detect mixed embedding-model versions, then rebuild or isolate incompatible vector partitions until queries and stored vectors use a compatible model.
Change only S3 storage encryption settings and leave the mixed embeddings searchable.
Increase the number of small index shards even though shard sizing is not the source of poor similarity results.
Lower the query similarity threshold without checking whether vectors from different embedding models are comparable.

Explanation

Correct Answer

Vectors from different embedding spaces are not safely interchangeable. Validate the version per partition and re-embed/migrate consistently.

Incorrect Answers

Encryption is required for protection but does not make two distinct embedding spaces semantically compatible.

This is a vector-version mismatch, not the oversharding and coordination overhead evaluated in another practice scenario.

Threshold tuning cannot repair incomparable embedding representations; it may instead increase false matches.

Question 96

An auditor needs a citation to the exact PDF page supporting a generated claim, but indexed chunks no longer retain page references. What evidence mapping must survive ingestion?

Store only the nearest-neighbor score and omit source-page mapping.
Preserve source document identity and page/location metadata through parsing, chunking and retrieval so citations can reference the actual page.
Reconstruct the page number by asking the FM to guess from the retrieved text.
Use no-chunking mode and assume Bedrock automatically provides its usual page-number metadata filter.

Explanation

Correct Answer

An answer can cite a source only as precisely as its stored provenance. Bedrock chunking options affect whether page-number metadata is retained; no-chunking configurations have documented limits.

Incorrect Answers

A similarity value says nothing about which PDF page contains the supporting passage.

A generated page label is not a reliable citation or a deterministic mapping to the document.

AWS explicitly documents that no-chunking mode does not provide the page-number citation/filter metadata field.

Question 97

A regional vector store is a single availability point, and the workload has a recovery-time objective requiring continued retrieval after instance/zone failure. What design assessment is needed?

Add more indexing dimensions to make storage survive an Availability Zone outage.
Increase query top-k while leaving the only index replica in one failure domain.
Rely on FM pretraining to reconstruct current private records during an outage.
Select a store-supported multi-AZ, replication or backup/recovery configuration and test failover against actual availability and recovery objectives.

Explanation

Correct Answer

Availability depends on the capabilities and placement of the chosen vector engine. Replication/recovery design must be tested rather than assumed from a single region label.

Incorrect Answers

Vector dimension controls embedding shape, not fault tolerance.

Retrieving more candidates cannot restore an unavailable index.

Model weights are not a replacement for an authoritative, current private vector corpus.

Question 98

An auditor asks why a particular vector hit was used to answer a disputed clause. The application currently records only the generated answer text. What retrieval evidence should be retained?

Save only the FM's final prose and discard all retrieved document identifiers.
Log the authorized source document/version ID, chunk ID, applicable retrieval score and query/run correlation for each cited hit.
Store only the vector similarity score without a document/chunk reference.
Assume two semantically similar retrieved passages are interchangeable evidence in an audit.

Explanation

Correct Answer

A reproducible retrieval trace identifies which version and chunk informed the answer, while preserving permissions and retention safeguards for the audit record.

Incorrect Answers

The resulting record cannot establish which source revision or chunk produced the disputed claim.

A numerical score cannot identify the document or its effective version.

Similar passages may have different approval states, dates and authoritative meaning.

Question 99

A Knowledge Base rebuilds every document nightly, exceeding its maintenance window even though only a small subset changes daily. Which ingestion strategy is supported?

Re-embed every unchanged object on each query before returning a result.
Disable source syncing entirely and rely on the model to remember policy edits.
Use incremental changed-document synchronization with durable checkpoints and validate additions, updates and deletions against source inventory.
Increase the retrieved-result count while keeping the nightly full-rebuild schedule.

Explanation

Correct Answer

Bedrock data-source synchronization can process changed, added and removed content incrementally. Full rebuilds waste embedding work and expand the update window unnecessarily.

Incorrect Answers

Query-time full reindexing is slower and more expensive than incremental ingestion.

Model weights do not automatically learn new source documents or deletions.

Top-k search settings do not reduce data-source processing volume or the time spent generating embeddings.

Question 100

An engineering team proposes replacing its vector store, but no measurements compare retrieval outcomes. What acceptance criteria justify the switch?

Select the store with the largest advertised vector count and skip retrieval testing.
Compare one store using easy English queries and the other using long multilingual queries.
Benchmark candidate stores on identical representative queries for authorized-access behavior, recall, tail latency, availability and total operating cost.
Choose whichever vendor provides the shortest free-trial period as proof of reliability.

Explanation

Correct Answer

A store decision is a workload evaluation, not a vendor popularity contest. Same-query testing must cover the security, retrieval-quality and reliability constraints the application needs.

Incorrect Answers

Capacity alone does not prove access control, useful recall or acceptable p95 latency.

Changing the evaluation population prevents a fair relevance and performance comparison.

Trial length is unrelated to availability and recovery characteristics under the intended production load.

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

Question 90

Explanation

Correct Answer

Incorrect Answers

Question 91

Explanation

Correct Answer

Incorrect Answers

Question 92

Explanation

Correct Answer

Incorrect Answers

Question 93

Explanation

Correct Answer

Incorrect Answers

Question 94

Explanation

Correct Answer

Incorrect Answers

Question 95

Explanation

Correct Answer

Incorrect Answers

Question 96

Explanation

Correct Answer

Incorrect Answers

Question 97

Explanation

Correct Answer

Incorrect Answers

Question 98

Explanation

Correct Answer

Incorrect Answers

Question 99

Explanation

Correct Answer

Incorrect Answers

Question 100

Explanation

Correct Answer

Incorrect Answers

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