Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 8

Question 70

A Bedrock assistant request inadvertently includes internal debug headers containing employee email addresses and access tokens. The business question needs neither field. Which request-assembly control is required?

Send all debug fields and rely on the model to refrain from repeating them in the answer.
Add a larger output safety filter without changing the input fields.
Apply purpose-based allowlisting and remove unnecessary sensitive metadata before constructing the model input.
Move the raw debug headers into an unfiltered system message and keep them for future support.

Explanation

Correct Answer

Data minimization should happen before inference, logs or retrieval; irrelevant credentials and personal identifiers must not be forwarded as generative context.

Incorrect Answers

Once submitted, secrets are already in the inference payload regardless of response wording.

Output filtering does not eliminate unnecessary sensitive material in the request or its associated telemetry.

An alternate prompt role is not a safe storage or need-to-know boundary for secrets.

Question 71

An OpenSearch vector index is configured for 1,024 dimensions but its new embedding model emits 1,536-component vectors. Inserts fail with a dimension error. What must be corrected?

Align the index vector-field dimension and mapping to the exact embedding-model output, rebuilding the index as required before loading compatible vectors.
Increase only the number of document chunks while retaining the 1,024-dimensional field.
Turn off metadata filtering and submit the same incompatible vector.
Change the score threshold to zero so every vector is accepted.

Explanation

Correct Answer

A k-NN vector field has a fixed dimension set by its mapping. A vector of the wrong length cannot be made valid by changing the query ranking threshold; model and index dimensions must agree.

Incorrect Answers

More chunks increase vector count but cannot allow a vector with 1,536 components into a 1,024-dimensional field.

Access filters operate on indexed metadata after vector validation, not on the field length.

Similarity thresholds govern retrieval result selection and do not override index schema validation.

Question 72

A knowledge base was embedded with model V1, but the search application now encodes queries using V2 while stored vectors remain V1. Results become semantically incoherent. Which cutover is correct?

Keep both embedding models mixed and normalize the cosine scores to the same 0–1 range.
Keep query and indexed-document embeddings in the same compatible model space by restoring V1 queries or re-embedding/reindexing the corpus with V2.
Rerank unrelated V1 results with V2 without retrieving compatible initial candidates.
Increase the ANN search depth without changing the query/document model mismatch.

Explanation

Correct Answer

Equal vector dimensions alone do not ensure two embedding models share the same geometric meaning. The corpus and queries must be generated by a matched embedding contract.

Incorrect Answers

Score normalization does not align two independently trained embedding spaces.

Reranking cannot guarantee correct candidates when the first-stage retrieval uses incomparable vectors.

More approximate neighbors in the wrong embedding space cannot reliably restore semantic alignment.

Question 73

A multi-tenant RAG assistant returns one company's confidential policy text when another company asks a semantically similar question. Where must authorization be enforced?

Trust a system prompt instructing the model not to reveal another company's retrieved chunks.
Encrypt all stored vectors while still querying every tenant's metadata unrestricted.
Remove citations from the answer but keep the same unauthorized source chunks in context.
Resolve the authenticated tenant and apply effective document ACL/tenant filters before any retrieved text enters the model context.

Explanation

Correct Answer

Vector similarity does not carry permission to see a result. A retrieval filter tied to authenticated identity and document rights must prevent cross-tenant chunks reaching the prompt.

Incorrect Answers

A prompt cannot undo an unauthorized retrieval or provide a reliable identity enforcement boundary.

Encryption at rest does not establish which tenant may decrypt or retrieve plaintext in the service.

Hiding citations cannot prevent the assistant from using private source content.

Question 74

A legal knowledge base retrieves a superseded regulation because its vector store tracks authors but not dates of effect. Which indexed evidence field must be added?

Sort all chunks by embedding cosine similarity alone, regardless of their publication or effective dates.
Use the model's pretraining cutoff date as the current regulation timestamp.
Store source approval and effective-date/version metadata, then filter queries to the current authorized regulation before generation.
Increase source chunk overlap without adding any effective-date metadata.

Explanation

Correct Answer

Semantic relevance cannot determine whether a regulation remains in force; explicit source time and approval fields are necessary to constrain grounding to applicable evidence.

Incorrect Answers

An obsolete clause can be highly similar to a query while no longer legally applicable.

A model training date is not a document-specific effective date or proof of approval.

Overlap may retain text context but cannot distinguish current from revoked regulations.

Question 75

After adjusting k-NN HNSW search parameters, an OpenSearch collection gets faster but misses relevant labeled documents. Which evaluation should determine a safe setting?

Accept the fastest setting because every request returns at least one document.
Measure only embedding storage bytes and ignore missed document IDs.
Benchmark ANN recall against labeled nearest-neighbor ground truth while measuring p95 latency as search parameters vary.
Increase the generated answer length and assume it repairs absent ANN candidates.

Explanation

Correct Answer

Approximate-nearest-neighbor settings trade recall for cost/latency. Labeled queries and a known relevant set reveal whether a faster setting dropped needed evidence.

Incorrect Answers

Nonempty results do not prove retrieval recall or relevance for the test corpus.

Index size does not quantify whether the expected relevant neighbors are found.

Generation cannot recover documents that the retriever failed to return.

Question 76

An architect proposes storing 384-dimensional vectors from one language embedding model and 1,024-dimensional vectors from an unrelated model in a single OpenSearch k-NN field. What index design must change before ingestion?

Set the common field dimension to 1,024 and pad the 384-dimensional values with zeros.
Create distinct vector fields or indexes with their own required dimensions and matched query embedding models, rather than one shared incompatible field.
Store both vector lengths as text metadata in a single 384-dimensional k-NN field.
Increase the result count until cross-model vectors become similar.

Explanation

Correct Answer

Correct: an OpenSearch vector field has a fixed dimension, and vectors from unrelated model spaces cannot share a meaningful similarity metric. Separate fields/indexes preserve each dimensional and semantic contract at design time.

Incorrect Answers

Zero-padding might make numeric lengths equal but does not make unrelated embedding coordinates comparable; it corrupts similarity semantics.

Metadata can record vector provenance but does not change the field's dimensional constraint or represent 1,024 values in 384 components.

Retrieving more candidates cannot make incompatible embedding spaces or dimensions valid for one vector field.

Question 77

An OpenSearch Serverless vector collection holding confidential KB data is reachable from arbitrary public IP addresses. Which change directly closes its network exposure?

Replace public collection access with a private network policy granting only approved VPC endpoints and necessary Bedrock service access, plus separate data permissions.
Add a restrictive data-access policy while leaving a collection-wide public network policy unchanged.
Disable the OpenSearch dashboard but leave the collection API endpoint publicly accessible.
Encrypt the index with a stronger KMS key while retaining unrestricted network ingress.

Explanation

Correct Answer

OpenSearch Serverless separates network policies from collection/index data access policies. A private collection policy must not include a matching public-access rule, which would take precedence.

Incorrect Answers

Data permissions are necessary but do not remove network reachability from the public internet.

Collection and dashboard endpoints have separate network access rules; hiding only Dashboards leaves the data API exposed.

At-rest encryption does not close a publicly reachable collection endpoint.

Question 78

A vector search must join product embeddings with existing tenant-owned relational purchase tables in Aurora PostgreSQL. Which storage design is worth evaluating?

Use an OpenSearch vector collection and assume it directly executes the existing PostgreSQL joins.
Move all tenant purchase rows into the model prompt and have it simulate the join.
Set the vector dimension to the maximum supported value without checking model compatibility or join requirements.
Evaluate a supported Aurora PostgreSQL pgvector deployment with appropriate vector indexes, relational joins and enforced tenant-level SQL authorization.

Explanation

Correct Answer

Aurora PostgreSQL supports pgvector subject to engine/extension prerequisites. Relational joins and permission controls can run alongside embedding similarity when carefully designed.

Incorrect Answers

OpenSearch is not a drop-in relational SQL join engine for Aurora business tables.

A probabilistic model is not an access-controlled relational executor and can produce unverified or leaking joins.

Dimension is an embedding-model contract, not a replacement for relational query semantics.

Question 79

A Bedrock Knowledge Base points to a pre-created OpenSearch Serverless vector collection but fails because the configured index fields do not match the stored schema. Which verification is required?

Grant the caller more Bedrock inference tokens without changing the OpenSearch field mapping.
Set an S3 object key to the name of an absent OpenSearch text field.
Match the vector, text and Bedrock-managed metadata field names/types to the knowledge-base storage mapping, including the chosen embedding dimension.
Turn off the knowledge base's source-citation feature to make an invalid index map acceptable.

Explanation

Correct Answer

Bedrock custom vector stores require designated vector, text and metadata fields; mapping names and compatible dimensions are specified when connecting the KB.

Incorrect Answers

Model output capacity cannot repair a missing storage field or wrong vector dimension.

An S3 source key is not an index field definition; schema must exist in the vector store.

Citations do not override required vector/text/metadata field names or index schema.

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.

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

Question 70

Explanation

Correct Answer

Incorrect Answers

Question 71

Explanation

Correct Answer

Incorrect Answers

Question 72

Explanation

Correct Answer

Incorrect Answers

Question 73

Explanation

Correct Answer

Incorrect Answers

Question 74

Explanation

Correct Answer

Incorrect Answers

Question 75

Explanation

Correct Answer

Incorrect Answers

Question 76

Explanation

Correct Answer

Incorrect Answers

Question 77

Explanation

Correct Answer

Incorrect Answers

Question 78

Explanation

Correct Answer

Incorrect Answers

Question 79

Explanation

Correct Answer

Incorrect Answers

Question 80

Explanation

Correct Answer

Incorrect Answers

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