Design a legal RAG assistant with trustworthy citations

knowledge, rag, citations, regulated

Prompt

A law firm wants an internal research assistant, grounded in a RAG pipeline over case law, that answers questions with citations. Early testing shows it sometimes cites a real, retrievable case that does not actually support the specific legal proposition the assistant attributes to it. Design a system that avoids shipping that failure to a filing or a client.

Interviewer's answers to clarifying questionsopen after you have asked your own

Corpus? Case law plus internal memos, continuously updated as new rulings and superseding guidance arrive.
Current check? "Is the citation a real, retrievable document," which passes even when the cited passage doesn't support the claim.
Use case severity? Output may end up quoted in a court filing or sent directly to a client; there is no tolerance for an unverified claim reaching either.
Staleness risk? Some guidance gets superseded; old and new versions sometimes sit in the corpus simultaneously without clear dating.
Review process wanted? A human lawyer reviews output before anything external goes out, but the firm wants the assistant to pre-filter obviously wrong citations.
Volume? ~400 research queries/week across the firm.

Follow-upsafter your high-level design

How would you build an automated check for citation-claim entailment, does the retrieved passage actually support the claim, not just citation existence?
What's your SLA for human review before any output reaches a filing, and what changes if that SLA is violated under deadline pressure?
The corpus contains both a superseded guideline and its replacement, with no date metadata distinguishing them. How does retrieval behave, and how should it?
How do you measure citation-claim accuracy in production without ground-truth labels for every query?

Answer first: out loud, on paper, or with the interview prompt above. Then compare with the rubric.