What Is RAG?

In plain language

RAG adds a retrieval step before an AI produces an answer. Rather than relying only on information already represented in the model, the system looks for material relevant to the question and supplies it to the model.

Google describes RAG as retrieving relevant data, adding it to the model input and generating an answer using that information. It is often used to make answers more accurate, timely and specific to the available material. (Google Cloud Documentation)

An everyday example

An analogy

RAG does not suddenly train the model on every retrieved document. It places relevant material in front of the model for the current answer. This is an editorial analogy, not a complete technical definition.

Where you will encounter it

You may see RAG in enterprise search, document question-answering, support assistants and knowledge-base products. It is relevant when answers should use a particular set of documents rather than only general model knowledge.

Should you care

• General chat users: knowing the basic idea is enough.
• Knowledge-base users: worth understanding.
• People building enterprise AI or search: essential.
• Teams evaluating answers: useful for asking which information was retrieved and supplied.

How it differs

RAG vs fine-tuning: RAG retrieves information for the current answer. Fine-tuning performs additional model training to shape behavior or patterns. They are not the same operation.

RAG vs Context: Retrieved information usually becomes part of the context available for the current response.

RAG vs Token: Tokens are the units a model processes. Retrieved material consumes tokens when it is added to the model input.

RAG vs API: RAG is a retrieve-then-generate approach. An API is an interface through which software communicates; a RAG system may use APIs as part of retrieval.

What to read next

Read What is AI context?, What is a token?, or What is an API?.