Housing staff worry the RAG bot cites lease clause IDs that do not match retrieved documents. What should they evaluate?

Whether RAG citations match the retrieved lease clause identifiers is what Housing staff should evaluate when the bot's references look wrong. Citation quality is a distinct grounding dimension: the assistant may retrieve the correct passage yet emit a wrong clause ID, or invent an ID that looks official in resident-facing text. Evaluating citation-identifier alignment against retrieval metadata catches misleading references that tenants and caseworkers might trust when disputing occupancy or lease terms.

Whether RAG citations match the retrieved lease clause content vectors checks internal embedding representations of passage text, not whether the cited clause identifier matches an identifier actually present in the retrieval payload. Whether RAG citations match the retrieved lease clause summary excerpts checks paraphrased content alignment, not whether the cited clause identifier matches what was actually retrieved. Whether RAG citations match the retrieved lease clause identifier ranks checks ordering position in the result list but still does not confirm that a cited clause identifier matches the clause that was actually retrieved.

Exam caveat: grounding evaluation for municipal RAG includes citation correctness relative to retrieved documents, not only fluent or helpful-sounding answer text. Operational check: for a labeled set of lease questions, compare cited clause IDs in the generative answer to identifiers present in the retrieval payload, and fail builds when the mismatch rate exceeds the housing authority's quality threshold.