From policy to practice: clause‑grounded answers with retrieval‑augmented generation

Abstract

Retrieval-augmented generation (RAG) supports policy question answering by grounding responses in institutional documents. This study describes a campus assistant developed with Azure OpenAI GPT-4 and an Azure Cosmos DB knowledge store containing university regulations. The evaluation covers 51 regulation documents organised into five categories and a question-answer set comprising 1,742 items. Human reviewers classified 1,644 system replies as correct, yielding an overall response accuracy of 94.37%. Response accuracy is an aggregate human-judged correctness measure, not exact match. Because the study did not retain item-level evaluation materials, it does not report separate measures of retrieval quality, citation fidelity, unsupported content, latency, or user experience.