A UML Profile for Domain-specific Modelling of Retrieval Augmented Generation (RAG) System Architecture

Retrieval-Augmented Generation (RAG) has emerged as the dominant architecture for mitigating hallucinations in Large Language Models (LLMs). However, the integration of probabilistic AI components with deterministic software introduces complex structural dependencies, leading to "silent failures" such as vector dimension mismatches and context window overflows. Current development practices remain largely code-centric and reactive, lacking rigorous design-time validation. To address this, this research applies Design Science Research (DSR) to develop a domain-specific UML Profile for RAG systems. The study formalizes a RAG Metamodel to define the abstract syntax of components like Chunker, Retriever, and implements Object Constraint Language (OCL) rules to enforce structural validity. The artifact is evaluated through scenario-based static analysis using defect injection. Results demonstrate that the proposed profile successfully detects critical integration defects specifically Data Dependency (TD1) and Interface Mismatches during the modelling phase, thereby enabling a "shift-left" approach to RAG reliability.