What is retrieval-augmented generation?

Definition

Retrieval-augmented generation (RAG) has two steps. First the system searches a set of documents (your policies, a client's file, a research library) for the passages most relevant to the question. Then it gives those passages to the language model and asks it to answer from them.

Done well, it means answers can cite the source and stay current without retraining the model.

Why it matters for advisors

RAG is how many advisor tools answer questions about your own content: "what did this client say about their daughter's tuition?" or "what does our policy say about texting?" It reduces hallucinations because the model works from real documents.

It also decides what the AI can see. If retrieval reaches every client's file, a staff member could pull information they shouldn't. Ask vendors how retrieval respects user permissions.

Good RAG tools show their sources. When an answer cites the exact policy paragraph or document page it came from, checking it takes seconds. If a tool can't show where an answer came from, treat it like any other unverified AI output.

Example

An internal assistant answers staff questions from the firm's compliance manual and shows the paragraph it used, so the reader can check it.

Watch out for

RAG lowers but doesn't remove errors. The model can still misread the passage it retrieved, so look at the cited source for anything important.