RAG Model

RAG stands for retrieval-augmented generation. A RAG system searches an external knowledge source for information relevant to a query and passes selected passages together with the question to a generative model. Documents are commonly segmented, indexed as embeddings, and retrieved by similarity, although lexical search or hybrid approaches can also be used.

RAG can provide current or internal information and support answers with source references without retraining the base model for every update. It does not guarantee correct answers. Unsuitable documents, missed results, poor segmentation, permission errors, or misinterpretation can still cause failures. Quality requires maintained sources, access control, retrieval and response evaluation, and traceable citations.

  • Combines retrieval with generative output
  • Uses external, updateable knowledge sources
  • Can use vector, lexical, or hybrid search
  • Does not eliminate hallucinations completely
  • Continuously evaluate sources, permissions, and retrieval quality

Frequently asked questions

A RAG system retrieves relevant information from external sources and provides it to a generative model as context for its answer.

No. RAG can improve grounding and verifiability, but retrieval, source, or interpretation errors may still produce false answers.
Zum Glossar