Privacy-Preserving Retrieval-Augmented Generation for Sensitive Enterprise Data: A Zero-Trust Governance Architecture

Abstract

The recent introduction of retrieval-augmented generation (RAG), in which a parametric language model is conditioned on passages fetched from an external, non-parametric memory at inference time, offers enterprises a plausible route to grounding generative text systems in proprietary knowledge without retraining. Yet the same architecture that makes RAG attractive for knowledge-intensive tasks also creates a new and largely unexamined attack surface: the retrieval index itself becomes a queryable proxy for whatever sensitive records it was built from, and the boundary between “model” and “data” that traditional access control and perimeter security were built to police effectively dissolves. This paper develops, through a design-science methodology, a zero-trust governance architecture for deploying retrieval-augmented generation over sensitive enterprise corpora — financial records, health information, legal files, and internal communications. Drawing on NIST Special Publication 800-207, the differential privacy and federated learning literatures, and established enterprise data governance frameworks, we propose an architecture in which every retrieval call, embedding operation, and generation request is treated as an untrusted transaction subject to continuous authorization, attribute-based access control at the passage level, and privacy-preserving transformation prior to reaching the language model. We validate the design conceptually through STRIDE-based threat modeling and an  enterprise scenario rather than a production deployment, since RAG systems were, at the time of writing, newly proposed and not yet operationally mature. The paper contributes a reference architecture, a mapping of zero-trust principles onto retrieval-specific controls, and a comparative analysis of privacy-enhancing technologies for retrieval pipelines, intended to guide both researchers and enterprise architects contemplating early adoption of this technology.