The first time I properly understood what Retrieval-Augmented Generation (RAG) was, I stopped reading and sat with it for a moment.
Because I realised I had been doing it since 1989.
Not with vectors, embedding models, or anything that would have recognised the word *computational*. But the structure was remarkably similar.
A user comes with a need. You do not answer from general knowledge. You go to a verified and curated collection, retrieve the most relevant information, and construct a response based on what the documents actually say.
If the collection is well organised, the answer is likely to be good. If the collection is poorly organised, the answer suffers.
The quality of the response depends heavily on the quality of the knowledge underneath it.
That is RAG.
And, in another form, that is what a trained reference librarian was doing at a well-run database terminal thirty years ago.
RAG stands for *Retrieval-Augmented Generation*. The term became widely established through research published in 2020, describing an approach that combines information retrieval with language generation to address a major weakness of large language models: their tendency to produce plausible but unsupported information.
Instead of asking the model to rely entirely on what it learned during training, RAG connects it to a specific knowledge store at the time of answering.
Retrieve first. Generate second.
The distinction matters.
A 2026 systematic review published in *Applied Sciences* examined RAG and large language model systems in enterprise knowledge management. Among its findings was the potential of these systems to reduce information silos by making knowledge stored across disconnected documents easier to retrieve and use.
That has a very familiar ring to anyone who has worked with institutional information.
And here is the point I want every information professional to hear clearly:
*RAG is only as good as the knowledge store it retrieves from.*
A poorly organised, inconsistently described, incompletely catalogued collection will produce poor retrieval. The model cannot retrieve what cannot be found. It cannot reliably distinguish authoritative material from questionable material when the collection provides no useful signals. And it cannot consistently surface the most relevant document when the metadata does not accurately describe what that document contains.
Consider two environments.
A DSpace repository with good metadata, controlled vocabularies, consistent subject headings, clear provenance, and properly described institutional research output gives a RAG system something dependable to work with.
A folder containing thousands of vaguely named, duplicated and unsorted PDFs does not.
Both contain information.
Only one is prepared for reliable retrieval.
That difference is where the information professional comes in.
The person who understands classification, metadata, authority control, provenance, preservation and information discovery is not merely supporting the RAG system.
*She is helping create the conditions that allow the RAG system to work.*
MEDLARS was helping users retrieve relevant biomedical information as early as 1964. We did not call it RAG then. We did not have the terminology, the models or the computing infrastructure we have today.
But the underlying idea — retrieve trusted information first, then help the user make sense of it — is not new.
We have a name for the modern version now.
Use it.
And, more importantly, make sure the knowledge underneath it is worth retrieving.