A housing authority needs the resident bot to cite current occupancy rules from policy PDFs rather than invent answers. Which generative pattern should they implement?

Retrieval-augmented generation over the housing policy PDF corpus grounds the resident bot in current occupancy rules instead of invented answers. RAG retrieves relevant passages from authoritative PDFs, then conditions generation on that context so citations and rule statements track published policy. For housing authorities, occupancy, income, and waitlist rules change; grounding reduces hallucination risk on high-stakes resident guidance.

Fine-tuned generation over the housing policy PDF corpus for grounded occupancy answers bakes policy text into model weights rather than retrieving current passages at query time, so it goes stale after PDF revisions and cannot cite the specific clause supporting each answer. Map-reduce summarization over the housing policy PDF corpus for grounded answers condenses documents into broad summaries rather than retrieving query-specific clauses for generation, so it loses the precise clause-level context needed for accurate occupancy answers and reliable citations. Generative query expansion over the housing policy PDF corpus for grounded answers only rewrites or broadens the query to aid retrieval, but it does not itself retrieve passages and condition generation on them, so alone it cannot produce cited, grounded occupancy answers.

Exam caveat: RAG is the standard generative pattern when answers must cite or follow specific municipal documents rather than general knowledge. Operational check: ask occupancy questions whose answers exist only in recent PDF revisions and verify responses include retrieved clause content; remove the index temporarily and confirm ungrounded answers fail the citation quality bar.