A model cannot tell you what it was never shown. Everything else here follows from that.

Retrieval-augmented generation exists because a language model knows only what it was trained on, cannot cite what it knows, cannot be told to forget, and has no access to your book, your policies, your customers or anything that happened after its training ended. RAG supplies the missing material at question time, so the model reasons over evidence it can point at rather than over a compressed memory it cannot. That is the whole idea, and everything else is engineering: how documents are parsed and split, how text becomes vectors, how those vectors are indexed and searched, how sparse and dense retrieval are combined, how results are reranked, how the window is assembled, and how the answer is grounded in and cited back to its sources. The family has grown well beyond the original shape, and this guide covers all of it — naive and vector RAG, advanced RAG with its pre-retrieval and post-retrieval optimisations, modular RAG, graph RAG and its local and global search modes, hybrid and fusion retrieval, multi-index and federated RAG, structured RAG over SQL and tables, multimodal RAG, hierarchical and parent-document retrieval, late interaction, contextual retrieval, hypothetical document embeddings, query decomposition, self-reflective variants including self-RAG, corrective RAG and adaptive routing, active and iterative retrieval, long-context and cache-augmented generation, memory-augmented and personalised RAG, and then agentic RAG and multi-agent RAG, where retrieval stops being a fixed pipeline step and becomes a decision the agent makes, revises and repeats. It closes with evaluation, entitlement and security, and the operating model. Five lenses per concept: what it requires, how to architect it for a bank, a flow diagram, a plain-English reading, and what it hands off next.