Short answer
RAG retrieves relevant documents and feeds them to a language model so answers can cite fresher, private, or domain-specific evidence.
Why it matters
RAG is the default pattern for enterprise Q&A. Quality hinges on chunking, embeddings, ranking, and evaluation—not only the LLM brand.
Diagram
Article
Why
So answers can mention your policies—not only the model's old training memory.
Examples
Internal policy assistant
An HR chatbot retrieves the latest parental-leave policy chunk and quotes eligibility rules with a link to the source page. If no policy matches, it says it cannot find an approved document instead of inventing one.
Literature helper
A research tool retrieves abstracts related to a question and summarizes agreements/disagreements with paper links. The user still opens primary papers before citing them in academic work.
FAQ
Is RAG the same as fine-tuning?
No. Fine-tuning updates model weights using training data. RAG keeps the model mostly fixed and supplies external evidence at inference time. They solve different problems and are often used together.
What makes a RAG system fail?
Common failures include poor chunking, weak retrieval, missing access controls, stale indexes, prompt instructions that encourage guessing, and UI that hides sources. Evaluate retrieval and generation separately.
When should I use web search instead of a private index?
Use web search when freshness and public knowledge matter. Use a private index when answers must come from internal, permissioned, or licensed corpora. Some products combine both with clear provenance.
Sources
-
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis et al. / arXivFoundational academic paper introducing the RAG framing for knowledge-intensive NLP.
-
Anthropic documentation hub
AnthropicProvider documentation hub for building grounded assistants with modern LLMs.
-
Google AI developer docs
GoogleGoogle AI developer documentation for generative models and grounding-related product features.