RAG (retrieval-augmented generation)
Having an AI answer from documents you choose, by finding the relevant passages first and handing them to the model with the question.
Updated October 2026 · with examples from our own builds
What does RAG mean in practice?
On its own, a model answers from what it absorbed in training, which knows nothing about your business and may be out of date. With retrieval-augmented generation, the system first finds the passages that matter, such as help articles, a catalogue entry or a customer's own published claims, and asks the model to answer from those.
This is how a chatbot or assistant becomes specific to your product. Answer quality depends more on the retrieval and the source material than on the model, and outdated or thin documents produce confident, wrong answers.
Good systems also say when the sources don't cover a question, rather than letting the model fill the gap from memory.
RAG at a glance
- Answers from
- Documents and records you control
- Quality depends on
- The retrieval and the source material
- Should admit
- When the sources don't cover the question
Where has RAG come up in our builds?
The AI agent answers from the business's own help articles and documents rather than only from scripted replies, inside a persona and guardrails the business sets.
Herb, the plant assistant, answers from fact cards generated from the species catalogue and says so when a fact is missing.
For each sentence that makes a claim about the customer, code retrieves up to six candidates, the claims from the customer's own pages that share the most distinctive words with it plus everything the customer told us directly, and the model may cite only those.
Common questions about RAG
Can a chatbot answer only from my own documents?
It can be built to try. Retrieval supplies the right passages, instructions keep it to them, and a fallback answer or a handover covers questions the documents don't answer. Test it with questions your documents deliberately leave out.
What makes RAG answers go wrong?
Usually the sources: missing, outdated or contradictory documents, or retrieval that picks the wrong passage. Fixing the material often helps more than changing the model.
Related terms
Sources
Planning a product and not sure where RAG fits? Ask us, or browse all 116 terms in the glossary.