RAG
Retrieval-augmented generation — pulling relevant documents from a knowledge base before generating, so the LLM grounds its answer in your data.
RAG is the standard pattern for grounding an LLM in specific knowledge it wasn't trained on. You index your documents, retrieve the top-K most relevant chunks for the user's query, and prepend them to the prompt.
In agents, RAG shows up wherever the agent needs to read internal docs — customer support agents grounding on help-center content, research agents pulling from a private corpus, sales agents reading account histories.
The 2026 frontier is the move from naive top-K to hybrid retrieval (lexical + semantic + reranker) plus iterative retrieval where the agent issues follow-up queries based on what it found.
Where this shows up
Frequently asked
When do I need RAG versus fine-tuning?+
RAG when the knowledge changes often or is too large to fit in context. Fine-tuning when you want to change behavior, style, or domain-specific reasoning patterns. Most production stacks use both.
Can an agent do RAG itself?+
Yes. Agentic RAG is the pattern where the agent decides when to retrieve, issues its own queries, and iterates. It outperforms naive top-K on multi-hop questions but costs more tokens.