Overview
Retrieval-augmented generation, or RAG, addresses a central limitation of generative models: useful knowledge may be missing, stale, or difficult to verify inside model parameters. RAG makes the evidence path explicit by connecting retrieval with generation.
What it is
A RAG system searches a collection for passages, records, or documents relevant to a query and supplies them to a language model as context. The generator can then produce an answer constrained by, or at least informed by, those retrieved sources.
How it works
Documents are collected, segmented, indexed, and represented for retrieval. A query selects candidates through lexical, vector, or hybrid search; reranking can improve relevance. The model receives selected context and generates a response, ideally with citations and checks for unsupported claims.
Key concepts
- Chunking and indexing
- Dense and sparse retrieval
- Reranking
- Context selection
- Grounded generation
- Attribution and faithfulness
Current research questions
- How can retrieval find the right evidence when queries and documents use different language?
- How should systems detect missing, conflicting, or low-quality sources?
- What evaluation separates retrieval errors from generation errors?
- How can citations be made complete, precise, and useful to readers?
Applications
- Research literature assistants
- Enterprise knowledge search
- Technical support
- Legal and policy document analysis
- Evidence-aware question answering
Relevant research papers
A live selection from the ScholarLens research index. Open any result to read its full paper page and use the Reading Assistant where available.
Keep exploring
Search the literature with your own question, or open a paper and read it closely with ScholarLens.