RAG
RAG
Also known as: retrieval augmented generation
Fetching relevant documents first and answering from them, rather than from memory alone.
A language model knows only up to its training cutoff and can invent plausible answers. RAG retrieves documents for each question and hands them to the model.
Two gains follow: fresh or internal material becomes usable, and answers can carry sources. That matters in fields like education or medicine where evidence is required.
Retrieval decides quality. Fetch the wrong document and the model will ground a wrong answer in it.
- PipelineChunk, embed, retrieve, generate.
- CheckSomeone has to open the cited source and confirm the claim is there.