Grounding
Constraining a language model's output to information retrieved from a specified source — a search index, a document set or a database —…
GEO AIRAG
Retrieval-augmented generation
An architecture where a system retrieves relevant documents or passages first and then has a language model generate an answer conditioned on them.
In full
A typical RAG pipeline chunks source documents, embeds them into a vector index, retrieves the closest chunks to a query (often combined with lexical search), reranks them, and passes the survivors to the model as context. It underpins most AI search products and enterprise assistants. Because retrieval happens at the chunk level rather than the page level, how content is segmented and how self-contained each section is directly affects whether it is retrieved and quoted correctly.
Example
A documentation assistant retrieves three 500-token sections from different pages and produces an answer citing each.
Related terms
Embedding
A numerical vector representing the meaning of text, an image or another item, positioned so that semantically similar items lie close…
Vector search
Retrieval by nearest-neighbour similarity between embedding vectors rather than by keyword matching, usually implemented with approximate…
Semantic chunking
Splitting documents along meaning boundaries — sections, topics, complete ideas — rather than at fixed character counts, to produce…
Chunk retrieval
Retrieving individual segments of a document rather than whole pages, which is how most AI search and RAG systems select what to feed a…
Where to read more
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks · Landmark Post · arxiv.org