Retrieval augmented generation (RAG) has become the dominant architectural pattern for ground‐ ing large language model outputs in enterprise data. Despite extensive published work on individual components (embedding models, vector databases, reranking strategies, generation prompting), there is comparatively limited synthesis of the architectural patterns that distinguish a working RAG prototype from a production grade system in high stakes enterprise domains. This paper proposes a pattern catalog drawn from leading the design and operation of multiple production RAG systems in a large enterprise human capital management platform, including a support agent assistant that re‐ duced average ticket resolution time by approximately fifty percent. We catalog twelve patterns across four categories. ingestion and chunking, retrieval and reranking, evaluation and observability, and governance and safety. For each pattern we describe the problem it addresses, the solution structure, the consequences and tradeoffs, and the failure modes we observed when the pattern was applied incorrectly or omitted. The catalog is intended for practitioners building or operating RAG systems in domains where output quality, traceability, and governance are non negotiable.
Prashanth Reddy Pasham (2026) studied this question.