RAG, Explained Simply

A language model learned from a huge mass of text up to a certain date, but it knows nothing about your contracts, your documentation, or this week's news. RAG fills that gap without retraining the model.

The principle: search first, answer second. See also artificial intelligence.

01 — The problem

The model doesn't know what it hasn't seen

Asking an AI for information specific to your organization invites it to make up a plausible-sounding answer. RAG gives it the material before it answers.

02 — How it works

Search, supply, answer

• Your documents are split into passages and indexed.
• For each question, the most relevant passages are retrieved.
• They're supplied to the model along with the question, with instructions to rely on them and cite them.
• The answer is grounded in your content, with verifiable sources.

03 — What determines quality

Everything happens before the model

• Clean, up-to-date documents.
• Sensible chunking (neither too coarse nor too fine).
• A search that actually surfaces the right passages.
• An instruction that requires citations and saying "I don't know" when the information is missing.

04 — Limitations

What RAG doesn't solve

If the answer isn't anywhere in your documents, RAG can't do anything about it. You also need to manage access rights (not revealing to one person what belongs to another) and keep the index up to date.

05 — Frequently asked questions

Frequently Asked Questions

RAG or fine-tuning?

RAG brings up-to-date, citable knowledge without retraining. Fine-tuning adjusts style or format. They're often combined.

Do I have to send my documents to a third party?

Not necessarily: depending on the architecture, the index and the model can be hosted on your own infrastructure. That's an important factor when choosing.

Does RAG eliminate hallucinations?

It reduces them a lot when the answer is in the documents and citations are required. It doesn't eliminate them entirely.

06 — Resources & links

Where to go, concretely