Retrieval-Augmented Generation (RAG) systems offer an efficient approach to integrating Large Language Models (LLMs) with external knowledge sources, but they often raise significant data privacy concerns, especially when relying on cloud-based services. This paper introduces FAIRD (Fair AI Research and Education), a novel on-premises RAG framework designed for secure environments that addresses privacy and security risks associated with cloud-based LLMs. FAIRD improves data protection by storing and processing sensitive information locally and within secure networks, mitigating vulnerabilities linked to third-party services. This paper presents a comprehensive analysis of the FAIRD pipeline architecture, focusing on its ability to reduce model hallucinations through structured information retrieval and context augmentation. The framework is evaluated through local deployment at the University of Applied Sciences Kaiserslautern, demonstrating its efficacy in maintaining data sovereignty and compliance with regulatory standards such as GDPR. Critical security challenges, including indirect prompt injection and reconstruction attacks, are addressed and minimized through design principles that restrict external access and enforce local control. The modular design of FAIRD supports scalable and privacy-preserving LLM applications, offering a robust foundation for agentic systems in research, education, and industries requiring secure AI deployment. This work bridges theoretical advances in RAG with practical implementation, highlighting pathways for future research in autonomous, explainable, and secure AI systems.