The short answer
Retrieval-augmented generation (RAG) is a technique where an AI system first retrieves relevant passages from a set of documents, a database or the web, then passes them to a language model so it can answer from that evidence. Anthropic's glossary says RAG improves accuracy and grounds responses in evidence. Gemini Notebook (formerly NotebookLM), Perplexity's cited answers and Mistral's Vibe Libraries all work this way. RAG adds knowledge without retraining the model.
What RAG is, in plain English
A large language model on its own answers from what it learned in training. That knowledge is frozen and does not include your company's policies, last week's prices or the contract on your desk. RAG fixes that by giving the model an open-book exam: before it answers, the system finds the most relevant pages and puts them in front of it.
Anthropic's glossary defines retrieval-augmented generation as a technique that combines information retrieval with language model generation to improve the accuracy and relevance of the generated text and to better ground the response in evidence. Google's Machine Learning Glossary calls it a common grounding technique that improves an LLM's output with knowledge retrieved after the model was trained. NIST's Generative AI Profile names RAG, alongside grounding, as a data source organisations should consider, and asks them to document any RAG they apply when they deploy generative AI.
How RAG works in four steps
Amazon Web Services describes the process in four steps, and it is the same in almost every tool.
- Prepare the knowledge. Documents are split into chunks and converted into numbers called embeddings, stored in a searchable index (often a vector database).
- Retrieve. Your question is converted the same way and matched against the index to find the most relevant chunks.
- Augment the prompt. Those chunks are added to your prompt inside the model's context window.
- Keep it current. The documents and their embeddings are updated over time, so answers change when the source changes, with no retraining.
Anthropic adds a practical caveat: RAG is only as good as the knowledge base and what is retrieved. If the right passage is not found, the model cannot use it.
Where you already use RAG, November 2026
| Tool | How it retrieves | Good for |
|---|---|---|
| Gemini Notebook (Google, renamed from NotebookLM on 16 July 2026) | Answers are grounded only in your notebook sources, with citations. Sources include PDFs, Word, Google Docs, Slides, Sheets, web URLs and public YouTube videos with captions. | Studying a set of reports, policies or interview notes |
| Perplexity | Searches the web and includes inline citations in every answer. With a Microsoft 365 work or school account, @Microsoft 365 searches SharePoint and OneDrive files on Pro, Max, Enterprise Pro and Enterprise Max. | Current research where you want to click through to sources |
| Vibe (Mistral AI, formerly Le Chat) | Libraries are persistent knowledge bases built from your documents and web pages; answers carry numbered footnotes linked to the source, and documents are stored in Mistral's European cloud. | Teams that want document answers kept in the EU |
| Gemini | Type @Gemini Notebook in a Gemini chat to add sources and answer from them. Off by default in the EEA. | Mixing notebook answers into everyday chats |
Project features, where you attach files to a workspace, are the everyday version of the same idea. Anthropic says Claude projects on paid plans (Pro, Max, Ultra, Team and Enterprise) switch to a RAG mode when project knowledge nears the context window limit, expanding capacity by up to 10 times. Free accounts do not get that mode. For picking a tool, see the best AI for research and Gemini Notebook vs Perplexity.
Why RAG matters at work
- Answers from your own material. Policies, product specs, past proposals and meeting notes become answerable without retraining a model.
- Fewer made-up answers. Grounding in retrieved text is one of the main ways to reduce hallucination, and AWS notes that RAG lets the model present information with source attribution you can check.
- Cheaper and easier to update than fine-tuning. AWS points out that adding data this way avoids the high cost of retraining. Change the document and the next answer changes.
- Access control. AWS also notes that teams can restrict sources by authorisation level, so people only get answers from documents they are allowed to see.
Try it: build a small RAG system in five minutes
- Open Gemini Notebook with a Google account and create a notebook.
- Add three to five documents on one topic, for example your company's travel, expenses and leave policies.
- Ask a question that needs two documents, such as "Can I claim a taxi to the airport on a Sunday, and does it count as working time?"
- Click the citations and check each one supports the sentence it is attached to.
- Ask something the documents do not cover. A good RAG answer says the sources do not contain it.
Only upload documents you are allowed to share with a cloud service. See how to keep your data private in AI tools.
RAG vs fine-tuning
People often confuse the two. RAG changes what the model can see for each question; fine-tuning changes the model itself by training it further. Use RAG for facts that change or that must be cited. Use fine-tuning, where it is still offered, for a consistent format, style or narrow task. Most teams need RAG, not fine-tuning.
Common misconceptions about RAG
- "RAG means the model learns my documents." It does not. The documents are looked up each time and the model is unchanged.
- "With RAG it cannot hallucinate." It can still misread a passage or answer from general knowledge. Tell it to use only the sources and check citations.
- "More documents is always better." Messy, duplicated or out-of-date files lead to the wrong passage being retrieved. Curate the set.
- "RAG is only for developers." Uploading files to a notebook or project is RAG with the plumbing hidden.
Related terms
- Embedding explains how text is turned into numbers for search by meaning.
- Fine-tuning is the main alternative way to specialise a model.
- Hallucination is the problem RAG most often reduces.
- MCP is a standard way for AI tools to reach live data sources.
- Deep research: RAG over the web, run as a multi-step agent.
Prompts to copy
Grounded answer with citations
Using only the attached sources, answer: [question]. Cite the source and section for every claim. If the sources disagree, show both views. If they do not cover the question, say so and suggest what document would.
You get: A cited answer that shows conflicts between documents and admits gaps.
Find the gaps in your knowledge base
Here are the questions our team asks most often: [list]. For each one, say whether the attached documents answer it fully, partly or not at all, and quote the relevant passage.
You get: A gap list showing which documents your team needs to write or update.