Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 17

Question 160

A policy repository stores long manuals; fixed-size chunks repeatedly separate headings from their requirements. Which design choice best matches the stated constraint?

  1. Use modality-appropriate extraction with normalized metadata and validation per source type.
  2. Use structure-aware chunking that preserves headings, section boundaries and nearby definitions.
  3. Apply tenant and document ACL filtering before evidence enters model context.
  4. Use table-aware parsing and preserve row, column and unit metadata.

Explanation

Correct Answer

B:

Semantic boundaries make retrieved evidence interpretable and reduce context loss around policy clauses. Test section-preserving document chunks before generation using source-version and permission fixtures; retrieval cannot recover missing evidence essential to section-preserving document chunks.

Incorrect Answers

A:

The option might improve format-aware ingestion design; that remains insufficient for section-preserving document chunks. The case is resolved by section-preserving document chunks, which depends on a design that will use structure-aware chunking that preserves headings, section boundaries and nearby definitions.

C:

The design concern is real but different: pre-context permission filtering. It does not impose the specific requirement for section-preserving document chunks; a suitable solution must use structure-aware chunking that preserves headings, section boundaries and nearby definitions and demonstrate the result.

D:

This addresses table-structure preservation during ingestion; however, the problem asks for section-preserving document chunks. Without the safeguard to use structure-aware chunking that preserves headings, section boundaries and nearby definitions, the original failure mode is still possible.

Question 161

A dataset consists of scanned signed forms and embedding search is inaccurate because no readable text is indexed. Which architecture adjustment is most appropriate?

  1. Evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests.
  2. Use modality-appropriate extraction with normalized metadata and validation per source type.
  3. Add OCR and extraction validation before chunking and embedding.
  4. Apply tenant and document ACL filtering before evidence enters model context.
  5. Implement change detection, versioned reindexing and freshness monitoring.

Explanation

Correct Answer

C:

The index can only retrieve represented content; poor upstream extraction cannot be cured by search tuning alone.

Incorrect Answers

A:

That is primarily a technique for language-aware index coverage. It lacks the explicit measure for reliable source extraction before indexing that is necessary here, namely a system that will add OCR and extraction validation before chunking and embedding.

B:

Improving format-aware ingestion design does not imply reliable source extraction before indexing. The selected design must add OCR and extraction validation before chunking and embedding, whereas this choice omits that direct protection or verification step.

D:

This is a valid architectural concern when optimizing pre-context permission filtering. Here it leaves reliable source extraction before indexing unaddressed, whereas the intended change must add OCR and extraction validation before chunking and embedding before the case can be considered resolved.

E:

This would make sense as an additional project activity for freshness of indexed regulated sources. The priority required here is reliable source extraction before indexing, so the workflow still needs a control that will add OCR and extraction validation before chunking and embedding.

Question 162

A compliance team revises a regulation weekly but retrieval still serves the previous version. Which approach should be recommended?

  1. Add OCR and extraction validation before chunking and embedding.
  2. Implement change detection, versioned reindexing and freshness monitoring.
  3. Index region, product and effective-date metadata and filter before similarity ranking.
  4. Use modality-appropriate extraction with normalized metadata and validation per source type.

Explanation

Correct Answer

B:

An explicit ingestion lifecycle keeps vector and metadata indexes synchronized with approved source revisions.

Incorrect Answers

A:

The control would be justified if reliable source extraction before indexing were the limiting factor. In this scenario the decisive target is freshness of indexed regulated sources; success requires an architecture that will implement change detection, versioned reindexing and freshness monitoring.

C:

This shifts attention toward metadata-constrained retrieval when the relevant decision is about freshness of indexed regulated sources. Unless the architecture can implement change detection, versioned reindexing and freshness monitoring, the failure described in the stem remains unresolved.

D:

For a workflow whose primary weakness was format-aware ingestion design, that proposal could be reasonable. Here the unmet obligation is freshness of indexed regulated sources, and that requires the system to implement change detection, versioned reindexing and freshness monitoring.

Question 163

A retrieval service confuses identical policy names belonging to different regions. What is the best production design response?

  1. Implement change detection, versioned reindexing and freshness monitoring.
  2. Index region, product and effective-date metadata and filter before similarity ranking.
  3. Apply tenant and document ACL filtering before evidence enters model context.
  4. Deduplicate source content and diversify retrieval candidates.

Explanation

Correct Answer

B:

Metadata constraints prevent semantically similar but legally inapplicable documents from dominating results. Test metadata-constrained retrieval before generation using source-version and permission fixtures; retrieval cannot recover missing evidence essential to metadata-constrained retrieval.

Incorrect Answers

A:

This addresses freshness of indexed regulated sources; however, the problem asks for metadata-constrained retrieval. Without the safeguard to index region, product and effective-date metadata and filter before similarity ranking, the original failure mode is still possible.

C:

There is a legitimate use case for pre-context permission filtering, but it does not supply the test or enforcement needed for metadata-constrained retrieval. The design needs to index region, product and effective-date metadata and filter before similarity ranking.

D:

Consider its actual operating scope: nonduplicative retrieval coverage. The requested assurance is metadata-constrained retrieval, which calls for an implementation that can index region, product and effective-date metadata and filter before similarity ranking, not merely for a related improvement.

Question 164

A knowledge index contains duplicate copies of one manual, crowding out other relevant sources. It also retains deleted source documents in search results for days because source deletion is not propagated into the vector index. Which TWO ingestion controls fit? Choose TWO.

  1. Apply tenant and document ACL filtering before evidence enters model context.
  2. Propagate source tombstones to affected indexed chunks and verify deletion completion before returning those chunks.
  3. Evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests.
  4. Deduplicate source content and diversify retrieval candidates.
  5. Use table-aware parsing and preserve row, column and unit metadata.

Explanation

Correct Answers

B:

A source deletion must invalidate its derived embeddings and search records, not merely remove the upstream file. A tracked purge process prevents retired information from continuing to appear as current evidence.

D:

Duplicate or near-identical manuals can fill the top retrieval positions without adding distinct evidence. Hash-based content deduplication and candidate diversification improve coverage; deleted-document propagation is a different ingestion-lifecycle requirement.

Incorrect Answers

A:

This addresses pre-context permission filtering; however, the problem asks for nonduplicative retrieval coverage and propagation of source deletions into derived indexes. Without the safeguard to propagate source tombstones to affected indexed chunks and verify deletion completion before returning those chunks, the original failure mode is still possible.

C:

By itself, this improves language-aware index coverage, but the case is not asking the architect to optimize that dimension. The requirement is nonduplicative retrieval coverage and propagation of source deletions into derived indexes, supported by an approach to propagate source tombstones to affected indexed chunks and verify deletion completion before returning those chunks.

E:

The option might improve table-structure preservation during ingestion; that remains insufficient for nonduplicative retrieval coverage and propagation of source deletions into derived indexes. The case is resolved by nonduplicative retrieval coverage and propagation of source deletions into derived indexes, which depends on a design that will propagate source tombstones to affected indexed chunks and verify deletion completion before returning those chunks.

Question 165

A financial report loses relationships between table headers and values when ingested as plain sentence chunks. Which solution best resolves this operational gap?

  1. Use modality-appropriate extraction with normalized metadata and validation per source type.
  2. Deduplicate source content and diversify retrieval candidates.
  3. Use structure-aware chunking that preserves headings, section boundaries and nearby definitions.
  4. Use table-aware parsing and preserve row, column and unit metadata.

Explanation

Correct Answer

D:

Numeric evidence requires structural context; semantically plausible fragments can give incorrect interpretations.

Incorrect Answers

A:

This addresses format-aware ingestion design; however, the problem asks for table-structure preservation during ingestion. Without the safeguard to use table-aware parsing and preserve row, column and unit metadata, the original failure mode is still possible.

B:

Although nonduplicative retrieval coverage can be valuable, it is not the acceptance condition in the question. The missing provision is table-structure preservation during ingestion, addressed by the control to use table-aware parsing and preserve row, column and unit metadata.

C:

It is aimed at a different outcome, section-preserving document chunks. No evidence is given that it provides table-structure preservation during ingestion; the needed mechanism is to use table-aware parsing and preserve row, column and unit metadata at the appropriate boundary.

Question 166

A RAG index contains sensitive customer documents and filters are applied only after retrieved text reaches the model. What design decision would best fit these facts?

  1. Add OCR and extraction validation before chunking and embedding.
  2. Apply tenant and document ACL filtering before evidence enters model context.
  3. Deduplicate source content and diversify retrieval candidates.
  4. Evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests.

Explanation

Correct Answer

B:

Post-generation redaction cannot undo unauthorized source exposure during retrieval. Test pre-context permission filtering before generation using source-version and permission fixtures; retrieval cannot recover missing evidence essential to pre-context permission filtering.

Incorrect Answers

A:

Consider its actual operating scope: reliable source extraction before indexing. The requested assurance is pre-context permission filtering, which calls for an implementation that can apply tenant and document ACL filtering before evidence enters model context, not merely for a related improvement.

C:

The proposed measure is aimed at nonduplicative retrieval coverage, not pre-context permission filtering. The architecture must instead apply tenant and document ACL filtering before evidence enters model context; a separate change is still needed for pre-context permission filtering.

D:

The relevant trade-off is not primarily language-aware index coverage. A successful fix needs pre-context permission filtering; otherwise pre-context permission filtering remains unverified even after this change.

Question 167

A citation points to a page that has since been overwritten, so reviewers cannot reconstruct the answer. Which intervention most directly addresses the failure?

  1. Apply tenant and document ACL filtering before evidence enters model context.
  2. Implement change detection, versioned reindexing and freshness monitoring.
  3. Deduplicate source content and diversify retrieval candidates.
  4. Store immutable document versions and chunk-to-source references for cited material.

Explanation

Correct Answer

D:

Versioned provenance permits a claim to be checked against the exact evidence presented at generation time. Test reproducible citation provenance before generation using source-version and permission fixtures; retrieval cannot recover missing evidence essential to reproducible citation provenance.

Incorrect Answers

A:

Its main effect concerns pre-context permission filtering. The case explicitly hinges on reproducible citation provenance; to resolve it reliably, the implementation has to store immutable document versions and chunk-to-source references for cited material, which this choice does not establish.

B:

The option might improve freshness of indexed regulated sources; that remains insufficient for reproducible citation provenance. The case is resolved by reproducible citation provenance, which depends on a design that will store immutable document versions and chunk-to-source references for cited material.

C:

The proposed action prioritizes nonduplicative retrieval coverage. That can coexist with the right solution, but it cannot replace reproducible citation provenance: the application must still store immutable document versions and chunk-to-source references for cited material.

Question 168

A multilingual knowledge base searches only English embedding vectors, missing region-specific source text. What should the solution architect prioritize?

  1. Add OCR and extraction validation before chunking and embedding.
  2. Apply tenant and document ACL filtering before evidence enters model context.
  3. Use modality-appropriate extraction with normalized metadata and validation per source type.
  4. Deduplicate source content and diversify retrieval candidates.
  5. Evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests.

Explanation

Correct Answer

E:

Retrieval representation must support the actual source languages and users' query vocabulary. Test language-aware index coverage before generation using source-version and permission fixtures; retrieval cannot recover missing evidence essential to language-aware index coverage.

Incorrect Answers

A:

It emphasizes reliable source extraction before indexing without closing the stated gap in language-aware index coverage. By contrast, the remedy to evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests acts on the condition that makes this scenario fail.

B:

That is primarily a technique for pre-context permission filtering. It lacks the explicit measure for language-aware index coverage that is necessary here, namely a system that will evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests.

C:

Improving format-aware ingestion design does not imply language-aware index coverage. The selected design must evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests, whereas this choice omits that direct protection or verification step.

D:

This recommendation targets nonduplicative retrieval coverage at a different point in the workflow. It leaves language-aware index coverage outside the design contract and does not require the team to evaluate multilingual embeddings or language-aware retrieval with locale-specific relevance tests.

Question 169

A repository contains PDFs, spreadsheets and tickets but all are ingested through the same plain-text parser. Which architecture adjustment is most appropriate?

  1. Use modality-appropriate extraction with normalized metadata and validation per source type.
  2. Add OCR and extraction validation before chunking and embedding.
  3. Use structure-aware chunking that preserves headings, section boundaries and nearby definitions.
  4. Use table-aware parsing and preserve row, column and unit metadata.

Explanation

Correct Answer

A:

Reliable RAG starts with preserving meaning from each format, not assuming uniform text structure. Test format-aware ingestion design before generation using source-version and permission fixtures; retrieval cannot recover missing evidence essential to format-aware ingestion design.

Incorrect Answers

B:

This addresses reliable source extraction before indexing; however, the problem asks for format-aware ingestion design. Without the safeguard to use modality-appropriate extraction with normalized metadata and validation per source type, the original failure mode is still possible.

C:

The option belongs to the broader work on section-preserving document chunks, but it cannot by itself deliver format-aware ingestion design. A decision on this case needs the team to use modality-appropriate extraction with normalized metadata and validation per source type and verify that outcome.

D:

This shifts attention toward table-structure preservation during ingestion when the relevant decision is about format-aware ingestion design. Unless the architecture can use modality-appropriate extraction with normalized metadata and validation per source type, the failure described in the stem remains unresolved.

Question 170

An employee asks for a specific error code present verbatim in one troubleshooting page; vector search misses it. Which next step is most defensible?

  1. Use hybrid lexical-semantic retrieval with reranking tested on both query families.
  2. Use semantic vector retrieval with an evaluated embedding model.
  3. Filter on effective jurisdiction and date before semantic ranking.
  4. Add lexical or hybrid search for exact identifiers alongside semantic retrieval.

Explanation

Correct Answer

D:

Sparse keyword matching handles rare codes that dense similarity can underweight. Build query cases that verify exact-token retrieval for identifiers and check whether the chosen method returns complete, applicable evidence rather than merely similar text.

Incorrect Answers

A:

The design concern is real but different: mixed identifier-and-semantic search. It does not impose the specific requirement for exact-token retrieval for identifiers; a suitable solution must add lexical or hybrid search for exact identifiers alongside semantic retrieval and demonstrate the result.

B:

This addresses paraphrase-aware document retrieval; however, the problem asks for exact-token retrieval for identifiers. Without the safeguard to add lexical or hybrid search for exact identifiers alongside semantic retrieval, the original failure mode is still possible.

C:

That choice would be responsive to applicability-first document filtering, but the reported defect is different. It leaves exact-token retrieval for identifiers unproved because there is no commitment to add lexical or hybrid search for exact identifiers alongside semantic retrieval.

Claude Certified Architect - Professional (CCAR-P) Practice Tests