Retrieval and Provenance

Lesson

Terms

RAG, Retrieval-augmented generation (RAG), Embedding, Vector search, Reranking

Sources

• arxiv.org (opens in a new tab)

Check

• What problem does retrieval-augmented generation (RAG) primarily address?
  • It replaces the model entirely with a searchable document store
  • Weights are incomplete and go out of date, so sources are retrieved
  • Models cannot produce any text without retrieved examples to copy
  • Models run too slowly, so retrieval is used to shorten response time
• A RAG system retrieves the wrong document and answers confidently from it. Where did it fail?
  • The user’s keyboard
  • Retrieval: it supplied the wrong evidence
  • Generation only; the model misread a good document
  • Nothing failed; it answered as designed
• An SI answer sounds authoritative and cites nothing. What is the key literacy question?
  • Is it long enough to be complete?
  • Was it fast?
  • Does it sound confident and fluent?
  • Where did this come from?