Retrieval-augmented generation: look things up first, then let the model answer. Documents are turned into embeddings, the relevant chunks are fetched for each question and pasted into the prompt, and the model writes with them in front of it. It became the default way to give a chatbot your company's files, and the lead author has said he would have chosen a nicer name.
The name is not an acronym anyone was proud of. Patrick Lewis, the paper's first author, has said in interviews that they would have picked something better had they known it would catch on. The record notes this is true of most of the field's vocabulary.
The record's summary: RAG is an open-book exam. Fine-tuning tries to make the model memorise the textbook; RAG leaves the textbook on the desk. The model still has to read the right page, which is where most of the engineering goes and most of the demos fall down.
The first RAG demo I saw, in 2023, answered a question about an HR policy by quoting the policy, with a page number. The second answered by quoting a different company's policy that had been in the same folder. Both were fluent. That was my education in retrieval.
Sighted in 2024 in a consultancy deck with the phrase 'enterprise RAG pipeline', accompanied by a diagram of boxes and arrows that the record recognised from the 2020 paper, redrawn in the client's colours. Fees were not shown.
What does it mean? Write it the way you would say it out loud.
