Retrieval-Augmented Generation for Natural-Language Access to Enterprise Operational Data

Abstract

Enterprise data warehouses hold answers to countless operational questions, but for most employees those answers remain locked behind SQL syntax and schema knowledge they were never trained to acquire. Fixed dashboards address recurring questions well, but every question outside a dashboard's anticipated scope routes back to a data team, creating delays that discourage exploratory questioning altogether. Retrieval-augmented generation offers a categorically different access model: language models paired with a retrieval step grounded in an organization's actual schema and data can answer novel operational questions in natural language, closing a gap that ungrounded models cannot close at all. This article examines the production-engineering decisions, retrieval architecture, context management, execution guardrails, governance enforcement, and continuous evaluation discipline that determine whether such a system is trustworthy enough for operational decision-making and argues that the performance ceiling of a conversational data agent is set as much by the underlying warehouse's architecture as by the language model itself. The analysis synthesizes recent empirical evidence on retrieval-augmented text-to-SQL generation with established governance and evaluation frameworks and finds that production readiness in this domain is fundamentally a systems-engineering achievement rather than a model-selection decision.