Abstract

Organizations store decision-relevant information across financial reports, sales records, customer feedback, policy documents and market research, yet decision-makers often cannot retrieve the right facts quickly, and general-purpose AI chatbots answer from pre-trained knowledge that is neither current nor organization-specific. Retrieval-Augmented Generation (RAG) addresses this by retrieving relevant passages from a knowledge source before a large language model (LLM) generates its answer, which improves factual grounding and reduces hallucination. This paper reviews twelve studies spanning the retrieval-augmented generation lineage (RAG, REALM, RETRO, DPR, FiD, ColBERT), a recent survey of RAG techniques, a survey of LLM hallucination, and business-decision-support literature. It compares their methods and limitations, identifies gaps specific to enterprise deployment, comprising limited enterprise focus, multi-source data integration difficulty, residual hallucination, scalability and weak business-context awareness, and proposes a conceptual RAG-based Business Decision Assistant architecture: a query-processing layer, a dense retriever over an enterprise document index, an augmentation layer that fuses retrieved evidence into the prompt, and a generator that returns cited, confidence-flagged answers. The paper describes the architecture, the data flow, applicable governance and evaluation considerations, and a research agenda. The contribution is a literature-grounded design; it has not been implemented or benchmarked.

Keywords

How to Cite This Article

Rights