Retrieval-augmented generation (RAG)
Beginner
Letting a model look things up first, like checking a book, before it writes an answer.
Novice
A setup where a search step fetches relevant passages, from your documents or the web, and puts them in the model’s prompt, so it can answer from them and cite them. The documents can change without retraining.
Expert
Pairs the model’s parametric memory with a non-parametric index searched at query time. Beats fine-tuning for injecting facts, gives provenance, and is often combined with fine-tuning that teaches the model to use retrieved context.
See also: Fine-tuning, Hallucination.