Retrieval-Augmented Generation

Retrieval-augmented generation (RAG) retrieves relevant documents and conditions a generator on that evidence to improve factual grounding.

Look up useful docs first, then ask the model to answer using them.

RAG pipelines retrieve candidate passages (lexical, dense, or hybrid), optionally rerank them, and condition a generator on the selected evidence. Quality depends on chunking, indexing, permissions, and citation UX—not only the LLM.

RAG is the dominant enterprise pattern for grounding assistants in private knowledge without always fine-tuning.

• Internal knowledge assistants
• Customer support with knowledge bases
• Research Q&A over document corpora

• Improves factual grounding when retrieval is good
• Updates knowledge by reindexing docs
• Supports citations for EEAT

• Garbage retrieval produces garbage answers
• Access control must be enforced at retrieve time
• Does not eliminate all hallucinations

• RAG is not a single product—it is an architecture
• Adding a vector DB alone is not a complete RAG system

When is fine-tuning better than RAG?

Fine-tuning can help style, format, or specialized behavior; RAG is usually better for frequently changing factual corpora. Many systems combine both.

• Lewis et al. — Retrieval-Augmented Generation (arXiv)

• Lewis et al., 2020