What is RAG (retrieval-augmented generation)?

How does it work?

A language model like ChatGPT or Claude is trained on general data and doesn't "know" your business. RAG solves that in four steps:

Question: the user asks something.
Retrieval: the system finds the most relevant passages in your sources, typically through a vector store that searches by meaning.
Augmentation: those passages are passed to the language model together with the question.
Generation: the model forms an answer based on exactly those passages, and can point to the source.

Without RAG, a language model guesses. With RAG, it answers based on your real information.

What is it for?

RAG is the standard approach whenever an AI has to answer based on information that wasn't in its training:

Internal assistants: "What did we agree with client X?" answered from emails, documents and CRM.
Customer and website chat: visitors ask questions, the assistant answers based on your real content.
Support and knowledge sharing: answers from manuals, procedures or a knowledge base, with source references.

Why do you need it?

A language model on its own has two problems for business use: it doesn't know your specific information, and it can make up convincing-sounding nonsense (a hallucination). RAG tackles both: at the moment of the question the model is handed your current, correct facts, and bases its answer on them.

The alternative, fully retraining a model on your data (fine-tuning), is expensive, slow and has to be repeated every time your information changes. With RAG you simply update the sources: new info is available immediately, without retraining the model.

How Sevendays uses this: our website assistant and internal assistants run on RAG, on top of a vector store with your content and business data on European infrastructure. That's how our AI platform brings all your sources together and the assistant answers accurately, with source references. Our examples show concrete use cases, and the 5 stages of AI explains where this fits in your broader AI approach.

RAG or fine-tuning?

In short: fine-tuning adjusts the model's behaviour itself (tone, style, fixed tasks). RAG feeds the model current knowledge at the moment of the question. For business information that changes regularly, RAG is usually the better and cheaper choice. The two aren't mutually exclusive and are sometimes combined.

Related terms

Vector store: the database RAG retrieves the relevant passages from.
Embeddings: the meaning vectors that make retrieval possible.
Hallucination: a convincing but incorrect answer from a language model, exactly what RAG helps prevent.