CCAR-P - Integration (19% of the exam) - Section 3.5

Design a RAG pipeline with appropriate chunking and indexing strategies.

Building retrieval-augmented generation: how documents are chunked, what metadata travels with each chunk, how the index is built and refreshed, and how retrieved passages reach the prompt. Candidates should match chunk size and boundaries to the documents and recognise a stale or broken index as a quality failure.

chunking strategychunk metadataindex build and refreshembeddings

Practice question for this objective

Free sampleIntegrationhard

A university runs a student-advice assistant that answers questions from 3,200 programme handbooks and academic regulations. About 400 documents are revised each summer, and the team has been adding revised documents to the index without removing the versions they replace. In a sample of 600 answers, 21 percent quoted a rule that had been superseded and 9 percent quoted a rule belonging to a different programme. The registrar requires every answer to reflect the version in force for the asking student's own programme, and any revision to be searchable within one working day of its publication. A full rebuild of the index takes about 30 hours. Which two decisions BEST meet the requirement? Select TWO.

ARebuild the whole index from scratch every weekend, so that each Monday the stored chunks match the handbooks then published on the website.
BAttach programme code, academic year and document version to each chunk, and filter retrieval to the asking student's programme and current year. Correct
CAdd a system-prompt instruction telling the assistant to prefer the most recent version whenever retrieved passages give conflicting rules.
DOn each publication, re-chunk and re-embed only the revised document and delete its superseded chunks from the index in the same update. Correct
EMove to a larger embedding model so that revised handbook wording ranks above the superseded wording for the same student question.

Keep a RAG index current by refreshing changed documents incrementally and removing superseded chunks, and scope retrieval with chunk metadata such as version and audience. The failures come from the index, not the model: superseded versions remain retrievable, and nothing in the chunks says which programme or year they belong to. Deleting and replacing a document's chunks on publication removes the stale text within the stated window, and filtering on programme, year and version metadata restricts retrieval to passages that apply to the asking student before similarity ranking runs.

Why A is wrong: This is tempting because a clean full rebuild removes every superseded version in one pass. It is wrong because a revision published on a Monday would wait up to a week to become searchable, breaking the one-working-day requirement, and it does nothing to stop answers drawing on another programme's rules.

Why B is correct: This is correct because metadata that travels with each chunk lets the retriever exclude passages from other programmes and earlier years before ranking, which addresses both the cross-programme answers and the selection of the version in force for the asking student.

Why C is wrong: This is tempting because it is a one-line change that seems to resolve version conflicts. It is wrong because the passages carry no version or date the model can compare, so it cannot tell which is newer, and the instruction is a compensating control that leaves stale and wrong-programme chunks in the index to be retrieved.

Why D is correct: This is correct because an incremental refresh keyed on the document touches only the changed handbook, so it completes well inside one working day where a 30-hour rebuild cannot, and removing the old chunks in the same step means a superseded rule is no longer there to be retrieved.

Why E is wrong: This is tempting because better embeddings often improve retrieval quality. It is wrong because a revised rule and the rule it replaces are usually near-identical in wording, so no embedding model can reliably rank one above the other by similarity, and the stale chunks would still sit in the index.

See more CCAR-P practice questions, answers explained.

Exam traps in Integration

Answers that look right on this material and are not. Each one is a distractor from a different question in the CCAR-P bank for this domain.

The replacement policy was chunked differently, so its passages score lower than the 2019 passages for the same parking questions
Why it is wrong: Ranking differences between documents are a real retrieval issue and could explain which passage wins. It is wrong because the chunking settings are unchanged and, more fundamentally, a revoked document should not be in the index at all, so the question is why it is still there.
Add a system-prompt instruction naming the clinician's hospital and telling the assistant to disregard protocols from other sites
Why it is wrong: Tempting because it is quick and will reduce the error rate. It is wrong because the other site's passages still reach the prompt and the instruction is advisory, so a governance requirement for an enforced control is not met by relying on the model to ignore them.
Tighten the error-rate and p95 latency thresholds so smaller deviations page the on-call engineer, and add a new alert on retrieval index timeouts
Why it is wrong: This is tempting because it strengthens the monitoring the team already trusts and retrieval timeouts sound relevant to a retrieval problem. It is wrong because the regression produced no errors, no timeouts and no latency change: the feed change made the index return incomplete content quickly and successfully, so tighter operational thresholds would still have stayed silent.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.