One of the biggest limitations of a language model on its own is that it only knows what was in its training data — and that data has a cutoff date. Retrieval-augmented generation, or RAG, is the fix: instead of relying purely on memorized knowledge, the model is given the ability to look things up before answering.
How it works
A RAG system has two main parts. First, a retrieval step searches a knowledge base — this could be the live web, a company’s internal documents, or a specific database — and pulls back the most relevant snippets of text for the question being asked. Second, a generation step feeds those snippets to the language model along with the original question, so the model can write an answer grounded in that retrieved material rather than guessing from memory alone.
Why it’s a big deal
RAG solves two problems at once. It reduces hallucination, because the model has real source material to draw from instead of reconstructing facts from fuzzy memory. And it lets AI tools answer questions about information they were never trained on — your company’s internal wiki, this morning’s news, or a PDF you just uploaded — without retraining the entire model, which would be far too slow and expensive to do for every new piece of information.
Where you’ve probably already used it
Any AI assistant that can search the web and cite its sources, or that lets you upload a document and ask questions about it, is almost certainly using some form of RAG under the hood. It’s become one of the standard building blocks of practical AI products, precisely because raw language models alone aren’t enough for tasks that require current or private information.