Retrieval-Augmented Generation Enhanced with Feature Stores: A Hybrid Architecture for Enterprise Knowledge Management

This research paper presents a novel hybrid Retrieval-Augmented Generation (RAG) architecture enhanced with enterprise stores to enhance accuracy, personalization, and contextual relevance of knowledge retrieval in contemporary organizations. Unlike traditional RAG systems that rely solely on semantic similarity, the proposed approach addresses the critical limitations: context-unaware retrieval, limited personalization, and inconsistent result quality. To fill this gap, the paper combines real-time structured elements including user role, expertise, behavioral patterns, and document metadata to the retrieval process in a hybrid similarity scoring framework. This study employed a simulation-based experimental design encompassing 10,000 documents, 500 users and three distinct system configurations, which is the semantic-only baseline and two hybrid models with 10 and 50 user/context features. The experimental results demonstrate significant performance improvements in the hybrid models which include 15-20% higher Top-K retrieval accuracy, 36% improvement in Mean Reciprocal Rank (MRR) and 41% enhancement in user satisfaction metrics. Although the computational steps were added, the latency was still acceptable to the enterprise and even when the feature became stale the retrieval accuracy did not change. All in all, the results prove the integrative process between feature stores and RAG systems to be a strong direction to the more correct, personal, and context-varying enterprise knowledge management.