Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Short definition
An arrangement that, at every question, finds in a corpus of documents the passages closest to that question and has a model synthesize them to produce the answer.
Full definition
The acronym stands for retrieval-augmented generation. In its broad technical usage, it designates any architecture where a model fetches external elements before answering, which covers a documentary engine as much as access to a business database.
In the articles on this blog, the term is used in the sense of "classic RAG": you load raw documents, you ask questions, and the tool works on demand. NotebookLM, or a ChatGPT conversation with PDFs attached, are its common forms. The trait that counts isn't the retrieval technique but what subsists after the answer: nothing. The next question starts again from the same documents.
Usage in the field
Used as a term of comparison, to situate by contrast what a consolidated base brings: RAG designates the starting point you are trying to move beyond, not a technique to reject.
Synonyms and variants
"classic RAG" when you need to mark that you are talking about loading raw documents; "augmented document search" in a non-technical formulation.
Not to be confused with
- AI Wiki — a persistent base written and maintained by an agent, where each source updates the existing pages. It keeps what has been understood; RAG remakes it at every query.
- vector search — a method of matching by proximity of meaning between a question and texts. It is a component of RAG, and it serves just as well to query an already consolidated wiki.
- Session memory — a written trace, kept over the course of exchanges, of what was done, decided and left pending. It carries the unfolding of a piece of work; RAG carries only a documentary corpus.
Examples
A question requiring five documents to be cross-referenced: the arrangement finds passages in each, assembles them, answers — and starts over entirely at the next question.
Ambiguities / debates
The acronym is used sometimes for a precise software architecture, sometimes for the consumer usage that consists in attaching files to a conversation. The two overlap enough to be conflated, and the gap in meaning explains entire disagreements about what "RAG can't do".