Retrieval-Augmented Generation answers a question by first retrieving relevant passages from a knowledge source and then having a language model compose the answer from them. It solves two problems at once: a model cannot know facts outside its training data, and fine-tuning it on new facts is expensive and must be repeated whenever the facts change. It exists because most factual answers are cheaper to look up than to memorise. The cost is a new dependency chain - embeddings, a vector store, reranking and a generation step - which is why it is the most interconnected node in this graph.
Fine-tuning continues training an already pre-trained model on a smaller, task-specific dataset so that its weights shift toward the behaviour you want. It solves the problem that a general model may not follow a house style, a domain vocabulary or a fixed output format. It exists because pre-training is expensive but adaptation is not: starting from a pre-trained checkpoint means a modest dataset is enough. It is the main alternative to retrieval when the goal is behaviour rather than fresh facts.
