Concept
aka retrieval-augmented generation
RAG
Pattern where a model is grounded on retrieved documents fetched at inference time.
Definition
Retrieval-augmented generation pairs a language model with a search step: a query is embedded, top-k documents are retrieved from a vector database, and the model conditions its answer on them. RAG reduces hallucinations, enables fresh knowledge without retraining, and supports source citations.
Common use cases
- Enterprise search
- Customer support
- Documentation Q&A