Retrieval augmented generation (RAG)’s incorporation of knowledge from external databases is solving some of LLM’s persistent problems including hallucinations, outdated knowledge and untraceable reasoning processes. The architecture of this incorporation has taken three guises over recent years namely: Naive RAG, Advanced RAG and Modular RAG. Moving from native LLMs, to Naive, to Advanced to Modular RAG represents an evolution where each stage addresses the shortcomings of the previous generation. This piece explores how the architectures compare and how they can be implemented python.