RAG (Retrieval-Augmented Generation)

RAG stands for retrieval-augmented generation and describes a method in which a language model does not generate its answer from training knowledge alone, but first retrieves relevant information from an external source. In two steps: first the retrieval step, which finds the most relevant passages from a knowledge base for a question, then the generation step, in which the model writes an answer based on those passages.

Why RAG matters

A language model doesn’t know your prices, processes or policies — it wasn’t trained on them. Without RAG it fills such gaps with plausible inventions, so-called hallucinations. RAG ties the answer to concrete passages and reduces that risk considerably. At the same time, sources can be cited, making every statement verifiable — instead of something you simply have to believe.

RAG at Kyros

Kyros uses RAG to let assistants answer from your knowledge base — from crawled URLs, uploaded documents and hand-written golden answers. Grounded answers cite their sourceswith clickable footnotes, and if information is missing the assistant doesn’t guess but says so. For a deeper explanation, see the article RAG explained.