AI Sovereignty: Redundant Use of Large Language Models for Public Sector Resilience

Public institutions using generative Artificial Intelligence (AI) face risks like vendor lock-in, data breaches, and opacity, threatening digital sovereignty. This study introduces the Governance-Aware Retriever Framework (GnARF) to bolster institutional resilience. It integrates five core components: (1) Query Model Allocation (QMA) to reduce vendor dependency; (2) Response Extraction & Feedback (REF) for quality control; (3) Retrieval-Augmented Generation (RAG) to ground outputs; (4) Decision Logging for transparency and (5) a Personally Identifiable Information (PII) Filtering for privacy compliance. This modular framework enables public organizations to align AI deployment with democratic values, ensuring secure, auditable, and sovereign data management.