A Privacy-First Architecture for Fully Local Retrieval-Augmented Generation in Secure Document Intelligence

Large Language Models (LLMs) have accelerated the development of retrieval-augmented generation (RAG) techniques for information access and decision support. However, most existing RAG implementations rely on cloud-based inference or managed vector stores, limiting their use in privacy-sensitive or air-gapped environments. This work presents a fully local, privacy-preserving RAG architecture for document intelligence and question answering, which we refer to as RAGStack. The framework integrates PyMuPDF, FAISS, SentenceTransformers, and locally hosted LLMs via Ollama to enable end-to-end retrieval and generation without external network dependencies. We describe the system architecture, design choices, and operational trade-offs, and evaluate the framework through internal tests on CPU-only hardware. Results indicate that the approach delivers consistent retrieval grounding and practical latency for moderate-scale document corpora. RAGStack demonstrates the feasibility of offline RAG systems for domains with strict data-governance requirements and provides a reproducible reference implementation to support further research in private and auditable GenAI deployment.