Amazon AWS Certified Generative AI Developer - Professional AIP-C01 Practice Test 11

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.

Question 101

A long policy file is split every 1,000 characters mid-sentence. What retrieval mechanism or check should be selected for this case?

Increase Retrieve numberOfResults while leaving mid-sentence fixed character splits intact.
Change the similarity metric from cosine to dot product without rebuilding the chunks.
Rely on a larger generation model to infer missing text on either side of each split.
Adopt chunking that respects semantic/structural boundaries with measured retrieval quality.

Explanation

Correct Answer

Bedrock semantic chunking can honor shifts in subject and sentence boundaries; evaluate retrieval ground truth after replacing arbitrary character slices.

Incorrect Answers

More returned broken chunks cannot reliably restore sentence meaning or the missing clause boundary.

A distance metric changes ranking, not the boundaries at which ingestion broke policy clauses.

Generation cannot guarantee reconstruction of source clauses that the index never presents intact.

Question 102

A source PDF has very short clauses that require neighboring context. What should be done before treating retrieved content as authoritative evidence?

Tune chunk overlap to preserve context without excessive duplicated tokens.
Use zero overlap and lower only the generation temperature.
Increase the number of independent vector indexes but preserve the same missing clause context.
Append the entire PDF as a hidden system message to every query.

Explanation

Correct Answer

A tuned overlap includes neighboring clauses in adjacent text chunks and must be measured against redundant tokens and retrieval accuracy.

Incorrect Answers

Temperature controls sampling, not missing neighboring text from the retrieved evidence.

Index duplication does not repair context gaps at chunk boundaries.

This bypasses bounded retrieval, risks context overflow and does not evaluate chunk quality.

Question 103

A contract section title is separated from its clause body. Which correction most directly fixes this grounding failure?

Return only the highest-scoring child clause without its parent heading.
Ask the model to guess the missing section title from clause wording.
Use hierarchical chunking to retrieve the relevant clause with its parent heading.
Increase ANN recall with a wider vector probe but keep child-only citation context.

Explanation

Correct Answer

Hierarchical chunking retrieves a relevant child and substitutes the encompassing parent, restoring its governing contract heading.

Incorrect Answers

A clause without its scope and heading can be materially misinterpreted.

Model inference is not reliable evidence of the governing heading from the contract.

More child hits do not guarantee the parent heading is present with the cited clause.

Question 104

An appendix table is returned without column headers. What retrieval mechanism or check should be selected for this case?

Return the cell values as an unordered bag of text for a fluent model to interpret.
Use structure-preserving parser/table context or parent chunks for answer grounding.
Reembed the same headerless cell strings with a different model.
Apply a date-only metadata filter before generation but retain flattened table chunks.

Explanation

Correct Answer

A parser preserving table headers and row/column associations, or a parent context chunk, grounds each value in its original column.

Incorrect Answers

Without row-column associations, a value cannot be tied reliably to its column heading.

Better vectors cannot restore structural metadata discarded during document parsing.

Date filters cannot restore the missing header-to-value mapping needed to explain the table.

Question 105

A frequently updated price-report corpus is indexed for a customer assistant. Select TWO choices that avoid presenting model-pretraining guesses as a current price and obtain dated source evidence.

Re-rank old cached price records solely by semantic score regardless of effective date.
Give the model a longer prompt to infer the latest release from last month's documents.
Return a price from a page whose publication date is later than its source effective date without checking applicability.
Retrieve current source passages with effective dates.
Do not infer live pricing from static pretrained weights.

Explanation

Correct Answer

Retrieve dated, current source records and reject pretrained price guesses: a versioned index is evidence, unlike frozen model parameters.

The model's training corpus is not a live price feed; when updated sources are missing, abstain instead of inventing a price.

Incorrect Answers

Reranking cannot make an outdated price current.

No amount of context can prove a price absent from the available sources.

A page timestamp alone does not establish which price version is effective for the requested transaction.

Question 106

An embedding model cannot represent the needed multilingual text accurately. Which correction most directly fixes this grounding failure?

Translate only the query but leave source-language coverage unevaluated.
Evaluate compatible multilingual embedding performance on actual language pairs.
Switch vector distance metric without confirming the model supports both languages.
Rely exclusively on an English benchmark when the production documents are in Arabic and Urdu.

Explanation

Correct Answer

Embedding-model language coverage must be benchmarked on the actual source/query language pairs using labeled recall evidence and compatible vector spaces.

Incorrect Answers

An unverified translation path may still miss idioms, named entities and local terminology.

Changing score calculation cannot add missing multilingual semantic representations.

English accuracy does not establish cross-language retrieval quality.

Question 107

A retrieval index uses cosine but stored vectors differ in normalization. What retrieval mechanism or check should be selected for this case?

Validate similarity metric and vector preprocessing against model specifications.
Tweak only the chunk overlap without examining vector preprocessing.
Increase top-k to compensate for incompatible distance calculations.
Apply an effective-date filter as a replacement for validating the selected index metric.

Explanation

Correct Answer

Confirm the vector distance metric and model-specific normalization contract; compare query/corpus preprocessing with the index's similarity setting.

Incorrect Answers

Chunk overlap cannot correct inconsistent vector normalization.

More candidates cannot establish that similarity scores have the expected mathematical meaning.

Source recency and embedding distance compatibility are different tests.

Question 108

A dense-only retriever misses exact product IDs. What should be done before treating retrieved content as authoritative evidence?

Increase semantic-only top-k but never index or query the literal product ID field.
Enable a HYBRID override on a store without the required filterable text field.
Add supported hybrid keyword-and-vector retrieval for identifier-heavy queries.
Use an LLM to invent the closest-looking product ID when no search hit is found.

Explanation

Correct Answer

Bedrock HYBRID retrieval combines semantic vectors and filterable text where the chosen store supports it; test exact product-ID recall.

Incorrect Answers

Dense similarity may still miss opaque identifiers absent meaningful lexical similarity.

HYBRID is store/field dependent and may fall back to semantic search rather than solve exact-term retrieval.

Generated identifiers are not authoritative inventory evidence.

Question 109

A vector search gets relevant results but poor final rank. Which correction most directly fixes this grounding failure?

Enlarge the generation prompt without changing ordering of relevant passages.
Apply supported reranking and benchmark ranking quality before generation.
Change the vector-store encryption key as the ranking remedy.
Increase the candidate count indefinitely and omit a reranker quality benchmark.

Explanation

Correct Answer

Reranking the bounded retrieved candidates with a supported reranker directly addresses poor order, and must be measured on labeled relevance.

Incorrect Answers

The model will still see poorly prioritized evidence and may ignore the best passages.

Encryption protects data but does not repair the ordering of retrieved passages.

Extra low-quality hits add cost and context pressure without proving ranking improvement.

Question 110

A Knowledge Base returns snippets but the application also needs a narrative. What retrieval mechanism or check should be selected for this case?

Use RetrieveAndGenerate with source attribution, not Retrieve alone.
Call Retrieve only, then represent raw snippets as if the service wrote the narrative.
Use an unrelated Bedrock model response without passing the retrieved evidence or citations.
Insert a citation-looking URL in the final text without verifying retrieved spans.

Explanation

Correct Answer

RetrieveAndGenerate orchestrates retrieval and supported model generation, and can return source citations; Retrieve alone returns source hits.

Incorrect Answers

Retrieve returns evidence; it does not generate a composed answer.

A fluent model reply without the source passages is not a grounded narrative.

A synthetic link is not a provenance-backed citation to the actual Knowledge Base result.

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

Question 100

Explanation

Correct Answer

Incorrect Answers

Question 101

Explanation

Correct Answer

Incorrect Answers

Question 102

Explanation

Correct Answer

Incorrect Answers

Question 103

Explanation

Correct Answer

Incorrect Answers

Question 104

Explanation

Correct Answer

Incorrect Answers

Question 105

Explanation

Correct Answer

Incorrect Answers

Question 106

Explanation

Correct Answer

Incorrect Answers

Question 107

Explanation

Correct Answer

Incorrect Answers

Question 108

Explanation

Correct Answer

Incorrect Answers

Question 109

Explanation

Correct Answer

Incorrect Answers

Question 110

Explanation

Correct Answer

Incorrect Answers

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