Large Language Models have become a central component of modern artificial intelligence systems, but their use in knowledge-intensive applications remains limited by hallucination, outdated parametric knowledge, lack of provenance, and weak control over domain-specific information. Retrieval-Augmented Generation addresses these limitations by connecting language models to external knowledge sources at inference time. However, reliable RAG systems require more than simply retrieving documents and inserting them into a prompt: they depend on the careful design of ingestion, chunking, indexing, retrieval, reranking, context assembly, grounding, citation handling, security controls, and evaluation. This thesis presents a domain-agnostic design methodology for modular Retrieval- Augmented Generation systems and validates it through the implementation of an e-learning assistant.