P3 · 8 questions

Integration — RAG, MCP, Auth & Observability — Attempt 2

CCAR-P P3 practice. Some items are “select all that apply” — they tell you how many to pick and are graded all-or-nothing. Commit before revealing, then read every explanation.

0 of 8 answered

P3 · Q1MCP transport (stdio vs HTTP/SSE)
A developer is building an MCP server that runs on the same laptop as the host application and needs no network exposure. Which transport is the most appropriate default?
P3 · Q2Vector store selection & index freshness
A RAG system's source documents are updated hourly, but users complain the assistant cites last week's figures. The retrieval code and prompts are unchanged. What is the most likely cause?
P3 · Q3Contextual retrieval / query rewriting
In a multi-turn chat, a user asks 'What about the second one?' after a prior turn listed three pricing tiers. Retrieval returns irrelevant chunks. All fixes below are plausible; which most directly addresses the failure?
P3 · Q4Poor-RAG-answer triageSort each into a bucket
Sort each observed signal from a RAG quality review into the stage most likely responsible.
The correct source chunks are present in the context, but the final answer contradicts them.
Queries with exact part numbers return semantically-similar but wrong documents, and the right document never appears in top-k.
Retrieved passages routinely end mid-sentence and omit the definition that started in the previous block.
The model invents a policy detail that appears in none of the retrieved passages.
After migrating to a new embedding model without re-indexing old vectors, similarity matches degraded across the whole corpus.
P3 · Q5Rate limits / retries / idempotency
An agent bulk-calls a payment-status tool and occasionally hits HTTP 429 and transient 503s. A naive retry loop caused duplicate side effects during an incident. All options look reasonable; which is the least risky retry design?
P3 · Q6MCP server design / least privilege & tool descriptionsSelect 2
You are designing an MCP server that a model will call autonomously. Which TWO design choices most improve safe, correct tool routing?
P3 · Q7Observability spans / cost & latencySelect 2
A platform team wants production telemetry that lets them both debug individual agent runs and watch fleet-wide health. Which TWO signals should they capture per request?
P3 · Q8Enterprise identity & per-user data scoping
A RAG assistant is being rolled out across an enterprise that uses OIDC SSO. Different employees must retrieve from different subsets of a shared corpus based on their group memberships. Every option below is deployable; which is the least risky way to scope retrieval per user?

More P3 attempts