About The RAG Glossary
The RAG Glossary is a reference work: one page per term used in retrieval-augmented generation and the tooling around it.
Independent. No vendor affiliation, no affiliate links, no product rankings.
Why a glossary
Because the vocabulary of this field was assembled quickly, from three directions at once — information retrieval, machine learning, and vendor marketing — and the words don’t agree with each other.
Chunk has no standard definition; it means whatever unit a given pipeline retrieves. Index refers to a mathematical structure, a running service, and a billable resource, often in the same sentence. Document might be a file, a database row, a page, or a chunk, depending on which library’s documentation you’re reading. Semantic search is used both for a specific technique and as a synonym for anything involving embeddings.
None of that is a problem when everyone in the room shares an assumption. It becomes a problem when a term appears in a design doc, a paper, a job posting, or a contract, and the reader has to know precisely what was meant.
What an entry does
Anyone can paraphrase a definition. The parts worth writing are the ones that come after:
What it’s confused with. Every entry names its near neighbours and gives the actual distinction — embedding versus vector, index versus database, reranking versus retrieval, context window versus token limit. Most confusion in this field is between two things that are genuinely adjacent, and the fix is a sharp boundary rather than a longer explanation.
Where the usage is contested. When practitioners don’t agree on a word, the entry records the disagreement instead of quietly picking a side. That’s more useful than false precision, and it explains why two documents you’re reading seem to contradict each other.
Which layer is being discussed. The field routinely conflates the mathematical object, the implementation, and the product feature. Entries keep them apart and say which one a given usage refers to.
What this site does not do
No tutorials. No architecture advice. No “which should you use.” No benchmark numbers, invented or otherwise.
A glossary that starts recommending things stops being a glossary, and the format is the whole value here — the reason an entry can be linked to from anywhere as see: X is that its shape is predictable.
Accuracy
Figures that vary by model or library are given as hedged ranges, never as precise-sounding invented numbers. Named methods are described by their well-known behaviour; specific papers, authors, and published scores are not attributed unless genuinely known. Where something changes fast, the entry says so rather than freezing a snapshot into a definition.
Elsewhere in this network
For an explanation of RAG as a whole rather than its vocabulary — what-is-rag.com (Q&A) and retrieval-augmentedgeneration.com (guide). For measuring and debugging a system you’ve built — what-is-rag.net. For deciding whether to build one at all — what-is-rag.org.