Integration

Integration is the largest slice of CCAR-P and the least like Foundations. It tests whether you can wire Claude into systems that already have owners, schemas, latency budgets and auditors — not whether you can call an API.

Most candidates lose points by treating retrieval as a search-quality problem. The exam frames it as an architecture problem: which knowledge is authoritative, who is allowed to see it, how fresh it must be, and what you pay in tokens and milliseconds to bring it into context. Answers that bolt a vector database onto every question, or that solve permissions by filtering after retrieval, are the classic traps. So is assuming the agent's service identity can stand in for the user's.

The idea that unlocks the domain: context is a budget, and every integration decision spends it. Retrieval depth, tool discovery, caching, reranking, model tier and trace granularity all trade against the same accuracy · latency · cost triangle. Decide what the system must be right about, then buy that accuracy with the cheapest context that still survives an audit.

• Add retrieval when answers depend on private, large or fresh knowledge. If the corpus is small and stable, long context plus prompt caching beats a pipeline you have to operate.

• Long context wins on cross-document reasoning and rare-detail recall; retrieval wins on scale, freshness, per-query cost and citation granularity. Hybrids are normal — retrieve wide, then let a large window reason.

• Chunk on structure, not character count — section, clause, function, ticket. Overlap only to protect boundary sentences, and carry parent metadata on every chunk: title, path, source, date, ACL.

• Default to hybrid retrieval: keyword/BM25 for identifiers, codes and exact strings; embeddings for paraphrase; fuse the two lists, then rerank the top candidates down to the few you can afford.

• Reranking buys the most accuracy per added millisecond. Add a reranker before you reach for a larger generation model or a bigger top_k.

• Not everything is a document. Route counts, totals, joins and current-state lookups to the structured system through tool use; keep vector search for prose.

• Enforce permissions at query time inside the retrieval filter, with ACLs indexed alongside content. The model is not an authorization boundary.

• Propagate the end user's identity downstream via token exchange or on-behalf-of flows, so effective access is the intersection of user and agent scope and the audit log names the human.

• Least privilege per tool, not per agent: scope each connector to the narrowest read set and gate writes behind an explicit confirmation or approval step.

• Treat freshness as a design parameter — a staleness SLO per source, incremental versus scheduled full reindexing to meet it, and an as-of stamp on retrieved content so answers can flag age.

• Spend latency deliberately: stream to cut time-to-first-token, parallelize independent retrieval and tool calls, cache the stable prompt prefix, and route by difficulty — small tier for extraction and routing, reasoning tier for synthesis.

• Instrument end to end: one trace per request with spans for retrieval, each tool call and each model call, recording selected sources, tokens, latency and cost.

• Vector database reflex. Adding embeddings to a question that is really a structured query gives confident, unaggregatable answers. Ask what the source of truth is before choosing the retriever.

• Semantic-only search misses exact tokens — order IDs, error codes, SKUs, config keys. The tell is users saying the old keyword search worked better.

• Orphan chunks. A paragraph stripped of its heading answers confidently about the wrong product line or the superseded policy. Fix with metadata-carrying chunks, not a bigger top_k.

• Context stuffing. Extra low-relevance chunks dilute attention, add cost and introduce contradictions. Retrieve wide, rerank down.

• Permissions after the fact. Filtering results post-generation leaks; filtering after retrieval silently starves recall. Prompt instructions to withhold data are not access control.

• One shared service account for all users — the union of everyone's permissions, and an audit trail that only ever names the bot.

• Observability outside the compliance boundary. Prompts, retrieved passages and eval sets carry the same sensitivity as the source data: redact at the span, set retention, keep traces in-perimeter.

• Tool sprawl. Dozens of tool definitions preloaded on every request burn context and degrade selection. Prefer progressive discovery — standard connector protocols such as MCP make listing capabilities cheap, so preload the stable few and load the long tail on demand.

• Unbounded agent loops. No timeout, retry budget or circuit breaker turns one slow dependency into an outage. Degrade to a partial, cited answer instead.

• No golden set. Without labelled queries with known-correct sources you cannot tell a retrieval regression from a prompt regression after a model, embedding or index change — and re-embedding changes everything.

• Averages only. Enterprise SLOs live at p95/p99 and at time-to-first-token; the mean hides the cold cache, the reindex spike and the tail that users actually complain about.

• Unmeasured groundedness. Evaluate retrieval (recall, source correctness) separately from generation (faithfulness, citation accuracy). A right answer over wrong sources is luck.

BRIDGE — the order to reason about any Claude-to-enterprise integration.

• B — Boundary: name the system of record, the trust boundary and the compliance perimeter before picking any technology.

• R — Retrieve: choose the strategy — long context, hybrid search plus rerank, structured query via tools, or no retrieval at all.

• I — Identity: propagate the end user, filter by ACL at query time, scope each tool to least privilege.

• D — Deliver context: budget the window — preload the stable set, discover the long tail, cache the prefix.

• G — Gauge: set the accuracy · latency · cost target, then route by difficulty and stream to hit it.

• E — Evidence: trace every span, cite every claim, retain logs for audit and for evaluation.

Use it when a scenario connects Claude to enterprise data and asks for the first or best design move: the earliest letter still unresolved is almost always the answer.