RAG (Retrieval-Augmented Generation) is a method where a language model retrieves relevant material from an outside source before it answers, then writes its response based on what it just retrieved. It exists to fix two problems: models that invent facts from memory, and models stuck with stale training data that never saw the latest information or your private documents.
The One-Sentence Definition
RAG means “retrieve first, then generate.” When the model receives your question, it pulls relevant passages from a designated source, hands those passages to the model along with your question, and lets it write an answer grounded in that material. The retrieval step decides what facts the answer stands on; the generation step only handles the wording.
A Concrete Example
To see RAG in action, open Perplexity: you ask a question, it retrieves web pages, lists the sources next to the answer, and generates a cited response from them, which is the full retrieve-and-generate loop. Keep one thing in mind: a cited answer can still rely on a weak or unsuitable source, and the summary may omit the scope, caveats, or context of the original. That makes it excellent for building a source map quickly, but every important citation still deserves a direct look before you act on it.
How It Differs From Related Ideas
The clearest difference between RAG and fine-tuning is where the knowledge lives: fine-tuning bakes knowledge into the model’s weights, so changing that knowledge means retraining, while RAG feeds material to the model at answer time, so swapping in new material never touches the model itself. Compared with plain use, where the model answers straight from training memory, RAG adds the retrieval step, which lets the answer rest on sources you can actually check rather than on whatever the model happens to remember.
When You Actually Need It
You need RAG when your answer depends on material that changes, stays private, or has to come with a source, such as checking the latest policy, querying internal company documents, or writing research that requires citations. The “generation” half of RAG is just a language model; to feel how different open models answer with no retrieval to fall back on, try them one by one in HuggingChat. Because those answers may be inaccurate, you get the very reason RAG bolts extra material onto the model in the first place. Two boundaries to remember: the model list can change and some model access depends on the account plan, and for sensitive inputs or consequential outputs, retrieval still does not remove the need for human review.
