CCAR-P questions and answers — Domain 3: Integration

12 practice questions on Integration, the domain weighted at 19% of the CCAR-P exam. Every answer explains why the distractors fail — that is the part that transfers to questions you have not seen before.

Question 1

A European bank requires that inference on customer data occurs within the EU. Which consideration should drive the deployment decision?

1. Which surface has the newest features
2. Which surface offers regional endpoints guaranteeing inference within a named geography
3. Which surface is cheapest per token
4. Which SDK the team knows best

Answer: B. Data residency is a hard constraint, and cloud deployment surfaces offer regional endpoints precisely for it, typically at a premium. Feature currency, price and familiarity are real trade-offs but cannot override a requirement that determines whether the design is lawful.

Question 2

An agent queries a customer database using a single privileged service account. What is the architectural defect?

1. Service accounts are slower than user credentials
2. Any user can be returned any row, because authorisation is not enforced against the caller's identity
3. The credentials will expire
4. It prevents connection pooling

Answer: B. Authentication to Claude and authorisation over enterprise data are separate problems. The correct pattern propagates the caller's identity so the agent inherits that user's permissions, enforced at the data layer rather than trusted to the application.

Question 3

A RAG system returns confident answers that are not supported by the source corpus. Where should you look first?

1. The system prompt
2. What was actually retrieved for those queries
3. The model tier
4. The temperature setting

Answer: B. A large share of apparent hallucination in RAG is the model answering faithfully from chunks that never contained the answer. Check retrieval before touching prompt or tier — the fix is likely in chunking, contextual enrichment, hybrid search or reranking.

Question 4

Why does contextual enrichment improve retrieval?

1. It reduces the number of chunks stored
2. It restores the document context that chunking stripped away, so a chunk is interpretable and matchable on its own
3. It replaces the need for embeddings
4. It compresses chunks to save tokens

Answer: B. An isolated chunk such as "revenue grew 3% over the previous quarter" identifies no company and no period. Prepending a short generated description situating the chunk in its document makes it both retrievable and usable, and it improves embedding and lexical retrieval alike.

Question 5

Why combine embedding search with BM25 rather than using either alone?

1. To reduce storage costs
2. Embeddings capture meaning but miss exact identifiers and rare terms; lexical matching catches those but misses paraphrase
3. BM25 is faster, so it handles most queries
4. Embeddings are unreliable

Answer: B. The two methods fail in complementary ways, which is why hybrid retrieval outperforms either. Exact codes, part numbers and rare terminology are precisely what pure semantic search loses.

Question 6

What does adding a reranking stage contribute?

1. It replaces the need for hybrid search
2. It lets you retrieve broadly then keep only the highest-scoring chunks, improving relevance while shrinking the final prompt
3. It generates the answer
4. It removes the need for citations

Answer: B. Retrieve wide, rerank, keep the best. That raises accuracy and can lower cost and latency at the same time by cutting what reaches the model. Anthropic's published figures put contextual enrichment plus hybrid retrieval at roughly a 49% reduction in top-20 retrieval failure, and reranking on top at roughly 67%.

Question 7

Which is true of the Citations feature?

1. It works alongside strict structured output
2. It returns structured pointers into source documents, and cited_text does not count toward output tokens
3. It requires the model to be prompted to cite sources
4. It is only available for PDF documents

Answer: B. Citations return cited_text with a document index and a character, page or block location — guaranteed valid pointers rather than model-generated references. They are not compatible with strict structured output, because interleaved citation blocks conflict with a rigid schema, and they support plain text, PDF and custom content blocks.

Question 8

An enterprise wants one capability reusable by several internal applications and teams. Which integration mechanism fits?

1. Embed the logic in each application
2. Expose it as an MCP server
3. Give each team the raw API key and documentation
4. Build a separate agent per team

Answer: B. MCP standardises the interface so one server can serve any client, which is exactly the reuse-across-clients case. Duplicating logic per application creates drift, and distributing keys with documentation solves neither reuse nor governance.

Question 9

Your system delegates work to an external partner's agent. How should its responses be treated?

1. As trusted, since the partner is contracted
2. As untrusted input, subject to the same validation and injection defences as any external content
3. As tool definitions
4. As system-level instructions

Answer: B. A trust relationship with an organisation is not a trust relationship with the bytes crossing the boundary. Another system's output is external content to yours — validate it, and never let it issue instructions.

Question 10

Which set of signals should a production LLM integration log per request?

1. Errors and uptime only
2. request_id, model and prompt version, stop_reason, full usage, latency split, tool calls, retrieved versus cited sources, validation outcomes and cost attribution
3. Only the final response text
4. Request count per hour

Answer: B. The characteristic failure of an LLM system is a successful-looking response with wrong content, which error-and-uptime monitoring cannot see. The fuller set is what turns "quality seems worse" into a diagnosis, and it is also what an auditor asks for.

Question 11

A design adds reranking to a retrieval pipeline. What is the expected effect on the three-way trade?

1. Accuracy up, latency up, and cost potentially down because the final prompt is smaller
2. Accuracy up, latency down, cost up
3. All three improve
4. Accuracy unchanged, cost down

Answer: A. Reranking adds a scoring step, so latency rises. It improves relevance, and because it shrinks what reaches the model it can reduce generation cost. Recognising that a lever moves at least two of accuracy, latency and cost — usually in different directions — is the domain's core skill.

Question 12

Which is the correct order of latency levers for a user-facing feature?

1. Optimise latency first, then quality
2. Get quality right, then choose the lightest adequate tier, shorten the prompt and bound the output, stream so time-to-first-token governs the experience, and parallelise what is independent
3. Stream first, then reduce the model tier regardless of quality
4. Increase max_tokens so the model finishes sooner

Answer: B. The documented caution is to engineer for quality first, because optimising early hides what good looks like. After that the levers apply in roughly that order, and streaming matters because perceived responsiveness is governed by time-to-first-token rather than total time. Raising max_tokens does not speed anything up.

← Previous: Questions & answers — Domain 2  |  Next: Questions & answers — Domain 4 →