Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 9

Question 80

A Bedrock Knowledge Base can read its S3 documents but its ingestion run receives an access denial when attempting to write vectors to OpenSearch Serverless. Which permission path must be fixed?

Add S3 GetObject permissions to the calling end user but leave the KB service role's vector-write policy unchanged.
Rebuild document embeddings repeatedly without changing the denied principal/policy.
Enable a public OpenSearch endpoint to bypass missing index-level data permissions.
Grant the KB service role the narrowly required vector-store data-access permissions and verify the collection's network policy permits the Bedrock service.

Explanation

Correct Answer

Object-read permissions and OpenSearch Serverless index/write permissions are independent; the KB's execution role and applicable access policy must authorize the actual data-plane operation.

Incorrect Answers

The failure occurs when the KB service principal writes to the vector store, not when an end user reads the source object.

New vectors still fail to write if the service role is unauthorized.

Public network reachability cannot grant data-plane authorization and unnecessarily exposes the collection.

Question 81

Engineering and HR documents share a retrieval store, but employees should only query their department's approved corpus. What index architecture should enforce the separation?

Store every record under a single public index alias and ask the FM to select its department.
Use security-aware index separation or authenticated department metadata filtering before retrieval results enter model context.
Use separate cosine score thresholds for HR and engineering while keeping all results readable.
Encrypt all department chunks under the same accessible service role and skip per-document filters.

Explanation

Correct Answer

Index organization can encode department boundaries, but enforcement must derive from authenticated identity and document ACLs rather than asking the FM to suppress unauthorized results.

Incorrect Answers

The model can be given unauthorized chunks before generation, and prompt text is not a permission control.

Similarity thresholds do not restrict document authorization.

A broadly authorized caller can still retrieve restricted content from encrypted storage.

Question 82

A policy RAG index has author metadata but no approval state; draft and approved policies match queries equally well. What must retrieval enforce?

Select the highest cosine score regardless of policy approval.
Rename the draft documents to contain the word approved while leaving approval metadata absent.
Suppress the source link in the final answer and permit draft chunks to remain in context.
Index and filter document approval status and effective version so only approved, currently applicable policy chunks can be grounded.

Explanation

Correct Answer

Approval is a source-governance attribute, not a vector property. Adding/filtering it before generation prevents high-scoring draft text from being presented as official policy.

Incorrect Answers

Draft text may be semantically near-identical to approved policy but is not authorized source evidence.

Filename text is not an authoritative approval or effective-version control.

Withholding provenance does not change the unapproved status of the facts the model used.

Question 83

A source policy was replaced in S3 this morning, but Bedrock Knowledge Bases still retrieves the old clause. Select TWO actions that actually refresh the indexed text and verify removal of obsolete evidence.

Change only the chat UI label to the new policy effective date, without touching the data source.
Run the knowledge-base data-source sync after the new source document is available.
Increase the answer temperature while leaving the old indexed version untouched.
Retrieve against the affected clause and confirm obsolete chunks are no longer returned after synchronization.
Increase the search top-k so the obsolete clause stays eligible but is less visible.

Explanation

Correct Answer

Correct: syncing the data source processes added, modified or deleted source documents into the Knowledge Base.

Correct: post-sync evidence checks confirm that old indexed text and any relevant cache are no longer treated as current.

Incorrect Answers

A display label does not update the indexed document or remove stale embeddings.

Sampling settings cannot update the index or control which source version is retrieved.

More retrieval candidates cannot guarantee removal of a superseded policy from the searchable corpus.

Question 84

An S3 document has been deleted for compliance reasons, yet its old vector chunks still appear in answer retrieval. What correction addresses the deletion lifecycle?

Hide the source document from the user interface while leaving vectors and cache readable.
Replace the original S3 key with a new filename and keep all old chunks indexed.
Reduce model temperature so it does not quote the deleted record.
Synchronize source deletions into the vector index and invalidate any downstream cached retrieval results containing the removed document.

Explanation

Correct Answer

Knowledge-base source sync must propagate deletes; an application cache can still serve stale text until invalidation. Both boundaries must stop returning deleted evidence.

Incorrect Answers

Visual removal does not delete searchable embedding chunks or their stored text.

Changing the name does not eliminate existing vector hits, especially if their source linkage survives.

Sampling cannot reliably prevent retrieval or disclosure of still-accessible removed data.

Question 85

A vector store grew from thousands to millions of tenant records and p95 search time degraded. Retrieval recall must remain acceptable. Which physical-layout work is appropriate?

Double shards blindly without checking fan-out, shard overhead or recall.
Raise the model output-token limit to reduce vector query latency.
Benchmark partitioning/shard organization against tenant query patterns and p95 latency while verifying recall on representative queries.
Disable all tenant authorization filters to maximize query speed.

Explanation

Correct Answer

Index growth changes memory, fan-out and query distribution. A layout should reflect tenant/routing patterns and be accepted only with both latency and retrieval-recall evidence.

Incorrect Answers

Excessive shards may increase coordination cost and not fix the bottleneck.

Generation settings affect inference, not the vector store's search execution time.

Removing access controls is unacceptable and can mix unauthorized results even if latency improves.

Question 86

An OpenSearch vector collection has thousands of tiny, nearly empty shards following a growth experiment. All shard contents use the SAME embedding model; observed overhead is excessive relative to corpus size. What adjustment directly addresses this problem?

Benchmark and consolidate or resize the shard/index layout for measured corpus volume and query throughput.
Add policy effective-date metadata to the documents, leaving index topology unchanged.
Change the embedding field dimension without regenerating the stored vectors.
Rebuild vectors to correct a mixed-embedding-model failure, even though the scenario confirms all vectors use the same model.

Explanation

Correct Answer

Oversharding can increase coordination and metadata overhead. Right-size the index and measure retrieval latency and recall on the real workload.

Incorrect Answers

Effective-date filtering addresses document authority, not excessive shard allocation or search overhead.

Changing mapped dimensions arbitrarily would create compatibility problems rather than reduce shard overhead.

Embedding-model inconsistency is a different failure. Re-embedding uniform vectors does not fix an overfragmented index layout.

Question 87

A batch ingestion process embeds hundreds of large PDFs concurrently, repeatedly hitting throttling and memory limits. What throughput change preserves resumability?

Use bounded-concurrency ingestion with staged document batches and durable checkpoints so failures resume without reprocessing the whole corpus.
Run unlimited parallel embeddings and disable error reporting to avoid backpressure.
Reindex every PDF from the beginning after any single record fails.
Ask the FM to shorten all PDF content without preserving which pages were omitted.

Explanation

Correct Answer

Embedding/KB ingestion is a throughput and idempotency problem. Limiting parallel work and checkpointing successes prevents overload and reduces wasted duplicate indexing.

Incorrect Answers

Concurrency without a bound worsens throttling and hides lost source documents.

Full restarts amplify load and duplicate work instead of using durable progress.

Arbitrary truncation loses source evidence and does not provide a resumable ingestion strategy.

Question 88

A KB ingestion run reports success, but the source inventory has 10,000 documents and the index contains fewer than expected. What audit proves actual ingestion coverage?

Count only how many times the API returned HTTP 200 during the run.
Reconcile the source-object inventory against indexed chunk/document identities and each ingestion job's failed-file and status records.
Measure the highest vector similarity score for one example query.
Suppress failed-file records and claim index coverage from the upload total.

Explanation

Correct Answer

An ingestion-job success signal alone does not guarantee every intended source record is searchable. Inventory and failure reconciliation discovers omissions and partial indexing.

Incorrect Answers

API transport responses do not establish per-document indexing success.

One good retrieval result cannot prove thousands of intended source objects were processed.

Uploaded objects can fail parsing, embedding or index writes; failures must be tracked individually.

Question 89

A regulated customer requires vector embeddings to stay in approved AWS Regions and be accessible only by named roles. Select TWO independent controls covering location and confidentiality/authorization.

Verify the selected vector store's actual storage/processing Region and residency configuration.
Enable public index endpoints for troubleshooting while assuming storage Region controls protect access.
Assume a model's permitted inference Region is automatically the vector-store storage Region.
Apply supported encryption and least-privilege index/object access controls for approved service identities.
Rely on embedding similarity scores to prevent unauthorized users from reading sensitive source chunks.

Explanation

Correct Answer

Correct: vector-store geography is an independent data-residency control; inference routing and S3 object location do not automatically pin it.

Correct: encryption protects stored data while identity-based data access controls determine which principals may retrieve it.

Incorrect Answers

Public reachability is a separate exposure; an approved geography does not make the data access appropriate.

The vector collection may be a separate service in a different Region, so the actual store location must be verified.

Similarity is a ranking algorithm, not an authorization policy or encryption mechanism.

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.

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

Question 80

Explanation

Correct Answer

Incorrect Answers

Question 81

Explanation

Correct Answer

Incorrect Answers

Question 82

Explanation

Correct Answer

Incorrect Answers

Question 83

Explanation

Correct Answer

Incorrect Answers

Question 84

Explanation

Correct Answer

Incorrect Answers

Question 85

Explanation

Correct Answer

Incorrect Answers

Question 86

Explanation

Correct Answer

Incorrect Answers

Question 87

Explanation

Correct Answer

Incorrect Answers

Question 88

Explanation

Correct Answer

Incorrect Answers

Question 89

Explanation

Correct Answer

Incorrect Answers

Question 90

Explanation

Correct Answer

Incorrect Answers

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