Anthropic CCAR-P (Claude Certified Architect - Professional) Practice Test 19

Question 180

A trace records tool start and end times but no parent-child relationships. Concurrent branches are difficult to interpret. What should be added?

  1. Separate timestamp lists without parent-child relationships
  2. Only a global average tool duration
  3. Causal links between the request and its nested operations
  4. A longer final-answer transcript

Explanation

Correct Answer

C:

Parentage identifies which operation triggered another and supports reconstruction of concurrent paths. Timestamps alone may not reveal the dependency structure.

Incorrect Answers

A:

Timestamps can suggest order but do not reliably establish causality among concurrent operations. Nested spans need explicit relationships to the originating request.

B:

An aggregate removes more transaction detail. It cannot reconstruct which nested call belongs to which branch of the request.

D:

Answer prose does not reliably encode the execution graph. Causality should be captured by instrumentation as operations occur.

Question 181

A production trace shows an agent repeatedly selecting among six overlapping account-search tools with different defaults. What review is most useful?

  1. Consolidate or clarify overlapping capabilities based on required use cases
  2. Increase the number of tool examples without checking their consistency
  3. Remove every account-search tool regardless of business need
  4. Add another search tool with all defaults combined

Explanation

Correct Answer

A:

An unnecessarily ambiguous tool surface can reduce selection reliability. Review should preserve distinct required operations while removing redundant or poorly differentiated interfaces.

Incorrect Answers

B:

Examples can help only if they reinforce a coherent interface. More conflicting demonstrations may amplify the existing selection problem.

C:

Eliminating the capability abandons the task. Pruning should target redundancy and confusion while retaining the required functionality.

D:

A broader ambiguous interface may worsen the confusion. The first task is to understand which distinctions are necessary and how they are communicated.

Question 182

Documents were embedded with model A, but queries now use incompatible model B embeddings against the old index. Retrieval quality collapses. What should be corrected?

  1. Use compatible query and document embeddings in the same space
  2. Sort vectors by their file creation date
  3. Regenerate only query vectors with a new embedding model
  4. Lower the model's response temperature

Explanation

Correct Answer

A:

Similarity search assumes comparable vector representations. Changing one side without rebuilding or compatible mapping can make distances meaningless for relevance.

Incorrect Answers

B:

Temporal ordering does not establish semantic compatibility. The index and query representation must be designed to support the same comparison.

C:

Changing the query representation alone can worsen the mismatch. Queries and indexed documents must use compatible representations before their similarity is meaningful.

D:

Generation randomness is downstream of the retrieval failure. It does not align the query vectors with the indexed document vectors.

Question 183

A paragraph's explanation spans two chunk boundaries and is often retrieved incompletely. What ingestion experiment is most directly relevant?

  1. Keep the chunks but delete the adjacent text
  2. Reduce chunk size further while retaining the same hard boundaries
  3. Test boundary-aware chunks or modest overlap
  4. Increase the number of unrelated documents retrieved

Explanation

Correct Answer

C:

Overlap or structure-aware boundaries can preserve local continuity. The resulting index should be evaluated for both improved recall and redundant content costs.

Incorrect Answers

A:

Removing neighboring evidence makes the incomplete explanation harder to reconstruct. It reduces the very context the task is missing.

B:

Smaller fragments can separate even more context around a boundary. The experiment should preserve semantic continuity rather than intensify the known fragmentation.

D:

More unrelated material does not reliably supply the missing adjacent passage. The experiment should target chunk construction or linked retrieval.

Question 184

The final model is accurate whenever the answer passage is supplied, but retrieval rarely includes that passage. Which component should be improved first?

  1. A stricter instruction to avoid hallucinations only
  2. The output formatting template
  3. Candidate retrieval coverage for the relevant evidence
  4. The reranker operating only on the same retrieved candidates

Explanation

Correct Answer

C:

Generation cannot use a passage it never receives. Improving the retriever's recall addresses the observed upstream limit on answer quality.

Incorrect Answers

A:

Such guidance can encourage abstention, but it does not retrieve the needed passage. The system still requires better evidence coverage.

B:

Formatting affects presentation but does not cause the missing evidence. The experiment already shows generation succeeds when the passage is available.

D:

Reordering the current candidates cannot recover a passage missing from that set. The stated bottleneck is getting relevant evidence into the candidate pool.

Question 185

A first-stage retriever returns many plausible passages, but only a few directly answer the question. What role should a reranker play?

  1. Guarantee recovery of documents never retrieved
  2. Replace authorization checks with relevance scores
  3. Reorder candidates by query-specific relevance before context selection
  4. Create authoritative facts absent from every candidate

Explanation

Correct Answer

C:

Reranking can improve precision among already retrieved candidates. It should be evaluated separately from whether the initial retrieval found the needed evidence.

Incorrect Answers

A:

A reranker cannot select a document outside its candidate set. The first-stage recall still limits what can reach the final context.

B:

Relevance does not confer access. Candidate passages must remain subject to the requester's document permissions.

D:

A reranker scores supplied candidates; it does not establish new source facts. Missing evidence requires another retrieval or an explicit limitation.

Question 186

Increasing retrieved passages from ten to fifty improves recall but makes answers less focused and slower. What should guide the next configuration?

  1. Evaluate the recall, noise, latency, and answer-quality trade-off
  2. Assume a longer context window eliminates retrieval noise
  3. Compare only the number of returned documents
  4. Always choose one passage to minimize tokens
  5. Always choose the largest k because recall increased

Explanation

Correct Answer

A:

More candidates can expose useful evidence while also adding distraction and cost. The best setting depends on measured end-to-end behavior for the workload.

Incorrect Answers

B:

Capacity does not guarantee attention to the right passages. A large window can still contain misleading or irrelevant material.

C:

Count is a configuration variable, not a success measure. It must be related to whether users receive correct, useful, timely answers.

D:

A one-passage rule may omit complementary evidence. The observed trade-off calls for evaluation rather than a universal minimum.

E:

Recall alone does not capture the effect of irrelevant context or processing time. The final task must still meet quality and service requirements.

Question 187

A query asks for the policy effective in March, but the retriever returns the newest policy from September. What capability is needed?

  1. Always sort documents by the most recent upload
  2. Filter or resolve policy applicability by the requested time
  3. Remove dates from indexed metadata
  4. Prefer the document with the latest modification timestamp

Explanation

Correct Answer

B:

The query is historical, so latest is not necessarily correct. Retrieval must represent effective dates and select the version applicable to the requested period.

Incorrect Answers

A:

Upload recency does not answer a historical applicability question. The correct version may have been superseded for current use but remain relevant to March.

C:

Deleting temporal information makes version selection harder. The missing capability depends on preserving and using those dates.

D:

A file modification date need not be its policy effective date. The requested historical applicability must be resolved using the policy's governing time information.

Question 188

A document is removed from the source repository, but its chunks remain searchable for weeks. What lifecycle control is missing?

  1. Filter deleted documents only in the source repository's search interface
  2. Change only the document's display title
  3. Propagate deletions to indexes and relevant caches
  4. Run an incremental indexing job that processes only additions

Explanation

Correct Answer

C:

A retrieval system needs update and deletion semantics as well as ingestion. Otherwise removed content remains available through derived copies.

Incorrect Answers

A:

A separate retrieval index can still serve its own stored chunks. A source-interface filter does not invalidate those downstream copies.

B:

A display-title update does not invalidate indexed content or cached passages. The stale chunks can remain retrievable after the source document is deleted.

D:

An add-only job cannot remove content that disappeared from the source. The synchronization contract must represent deletions as well as new documents.

Question 189

A new index is built gradually while users query it, so some answers mix old and new document versions. What publication design improves consistency?

  1. Let each query select random old and new shards
  2. Mark the partially rebuilt index as current before validation
  3. Build a versioned index and switch readers after validation
  4. Delete the old index before ingestion starts

Explanation

Correct Answer

C:

A controlled cutover lets queries use a coherent index generation. Validation can check completeness before the new version becomes the serving target.

Incorrect Answers

A:

Random mixing preserves inconsistent evidence across the query. The architecture needs a defined version boundary rather than arbitrary combination.

B:

A current-version label does not make a partially rebuilt corpus consistent. Readers should switch after the new index meets its completeness and validation criteria.

D:

This creates an avoidable availability gap and still does not validate the replacement. A complete candidate should be prepared before cutover.

Question 190

A PDF loader silently skips unreadable pages, yet marks the document fully indexed. What should the pipeline expose?

  1. Page-level failures and document completeness status
  2. A larger generation model for every later query
  3. A success flag whenever at least one page loads
  4. Only the number of successfully embedded chunks

Explanation

Correct Answer

A:

Silent omission makes missing evidence look like a complete corpus. Explicit ingestion status allows remediation and prevents unsupported completeness claims.

Incorrect Answers

B:

A model cannot retrieve pages that were never indexed. The upstream omission needs detection and repair in the ingestion process.

C:

Partial success is not full ingestion. The status should distinguish a usable subset from a complete representation of the document.

D:

Success counts do not reveal which expected pages were omitted. The pipeline needs a denominator and failure details for meaningful completeness.

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