Retrieval-augmented generation (RAG)
Also known as: RAG
In short
A method that supplies retrieved information to a generative model so its response can use relevant external material.
Retrieval-augmented generation (RAG) combines information retrieval with generated responses. The system retrieves relevant material from a data source and provides it as context for the model’s answer.
That source might be a document collection, a product knowledge base, or a search index. The retrieved context supplements the model’s existing knowledge; retrieving a document does not retrain the base model.
A practical example
Imagine a support assistant answering whether a bicycle carrier fits a particular vehicle. It retrieves the manufacturer’s compatibility table and the relevant installation notes before composing a response.
If the table names the vehicle but excludes a certain roof type, that exception matters. A useful answer preserves it and identifies the supporting material. This is an illustrative workflow, not proof that every assistant performs these checks.
Retrieval does not guarantee correctness
RAG can provide more relevant information, but the retrieval step can miss the right document. A selected document can also be outdated, incomplete, or interpreted incorrectly.
Evaluate both what the system retrieved and what it claimed. For catalog questions, check item identity, revision dates, and whether the answer applies to the exact variant. Displaying a source link makes review easier but does not establish that the answer accurately reflects it.