What is RAG? A plain-English guide for business leaders

What is RAG? The acronym in plain English

RAG stands for retrieval augmented generation. Retrieval means the AI first fetches matching information. Generation means it then writes its answer from it. The term was coined in 2020 by a research team led by Patrick Lewis.

A simple question shows why it matters: "What's the notice period in our framework agreement with this supplier?" A general AI assistant answers fluently and confidently. The answer is still a guess, because the model has never seen the contract. It knows what such contracts usually say, not what yours says.

A RAG system searches your contracts first, finds clause 14.2 and answers "90 days, in writing", with a link to exactly that clause. For a company, that's the whole difference, because the questions that matter at work are almost always specific: what does our policy say, what did we agree with this customer, what has applied since March?

In shortRAG means: search your documents first, then answer, then show the source. That way the AI knows your company knowledge instead of guessing it.

How does a RAG system work?

A RAG system works in four steps. The first happens once at set-up and after every change; the other three happen with every question.

The trick is in step two. The system shouldn't just find passages with the same words, but passages that answer the question. Someone asking about the "notice period" should also find the sentence "The agreement ends 90 days after written notice."

To do that, text is converted into embeddings: long lists of numbers in which passages with similar meaning sit close together. They're stored in a vector database. Searching then means finding the passages closest to the question.

A good RAG system also checks permissions along the way. It only finds what the person asking is allowed to see. And because the documents aren't trained into the model, new material is available at once and deleted material is gone at once.

In shortIngest, search, answer, cite. Embeddings find matching passages by meaning, not just by matching words.

RAG, fine-tuning or a general AI assistant?

If you want AI to use your own knowledge, you'll meet three approaches. They solve different problems:

For knowledge that changes and has to be checked, RAG is usually the right fit in a company. Fine-tuning can come later, for example for a consistent tone.

Which models does a RAG system use? RAG isn't a model of its own but a method. It works with almost any language model: large models from international providers, models from EU-based providers, or local models on your own servers. The choice depends on data protection, cost and quality. In RAGSuite you choose the model per project and can switch at any time.

In shortFor knowledge that changes and has to be checked, RAG is usually the right fit. RAG is a method and works with almost any model.

Why the source on every answer is what counts

Picture someone in finance asking about the notice period and getting a sentence back. The next thirty seconds decide whether the tool was worth it.

If the answer links to clause 14.2 of the actual contract, they open it, read it and act. The system saved them twenty minutes of searching, and they checked it in half a minute.

If there's nothing underneath, they're left with two bad options: verify everything themselves, in which case the tool saved nothing. Or don't verify, and act on a sentence nobody checked.

That isn't a matter of interface taste. It's the difference between a system that survives an audit and one that quietly creates risk. Germany's data protection conference says the same in its guidance on RAG: RAG can improve accuracy, but false statements can't be fully ruled out. The source turns an invisible error into a checkable one.

In shortThe source turns a claim into something you can check. Test every demo with a question whose answer you know.

What a RAG system can't do

Each of these gets oversold somewhere, so here it is plainly:

• It doesn't make the model smarter. RAG gives the model better material. Good search with a weak model gives accurate but clumsy answers. Poor search with a strong model gives well-written answers from the wrong document, and those are harder to catch.
• It doesn't tidy your documents. If three versions of a policy sit in the folder, the system may cite any of them. RAG makes content problems visible and specific. Solving them is up to you.
• It doesn't replace an owner. Someone has to decide what goes in, what comes out when a new version arrives, and who may see what. Without that person, quality drops from month two.
• It doesn't decide permissions for you. The moment a system can read a folder, "who can see what?" becomes a real question. Settle it deliberately, before it comes up in an audit.

That's why a test on your own documents is worth more than any demo on someone else's: only there do you see how well search and model work together for you.

In shortRAG doesn't make the model smarter, tidy your files or replace an owner. Test with your own documents.

What do you need to start?

A first test needs less than many expect, but three things have to be right:

With RAGSuite, a knowledge engine for your data, you build a RAG system without development work: create an account in RAGSuite Engine or install the Community Edition on your servers, connect sources such as SharePoint or Confluence, choose a model. Every answer names its source. What that costs per month, you can work out with the AI cost calculator.

How the platform is built is shown on the enterprise RAG platform page. Whether RAGSuite Engine or your own servers fit better is covered in the guide RAGSuite Engine or self-hosted?. And if your RAG system processes personal data, read when a DPIA for AI is required.

In shortThree sources with lots of questions, a place to run it that fits your data, and a person who owns the content.