An approach that retrieves relevant information and supplies it as context for a model to generate a response.
Retrieval-augmented generation, or RAG, combines an information-retrieval step with a model’s generation of an answer. The system finds potentially relevant material and makes it available to the model as context. That material might come from a document collection, a database or a web-search system.
RAG does not require a live web search. An application can retrieve from an index built earlier, and the information available may reflect an older version of a source. Retrieval also differs from training: supplying a passage for an answer does not mean the model’s underlying parameters have been updated with that passage.
The resulting answer still needs evaluation. The retrieval step may miss the right document, return an outdated passage or include information that does not answer the question. The model may then misinterpret what it receives. A RAG architecture supports access to source material; it does not make every answer current or correct.
Takeaway: If a business uses a document-based assistant, ask which sources it can retrieve, when they were refreshed and how access is controlled. The architecture name alone does not establish reliability.
Example
A distributor’s assistant retrieves a maintenance paragraph from its stored product manuals before answering a customer question. If the manual is for the wrong model or an obsolete revision, the answer can still be unsuitable. The system needs the correct source and a faithful explanation.
What to check
Test representative questions against the actual underlying documents. Include cases where the answer is missing or the sources disagree. Keep private information restricted rather than assuming retrieval automatically applies the right permissions.
For more detail, see Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.