Types of RAG

The tuned pipeline answers most questions well. But some questions need two documents joined together, some answers hide in tables, PDFs or images, and sometimes the right answer is “I don’t know”.

This chapter: Which kind of RAG fits which kind of problem, and what does each one cost?

You will learn

Naive and advanced RAG: one search, tuned or not
Agentic RAG: an agent that decides when to search again
GraphRAG, self-correcting RAG and multimodal RAG, each on a question that needs it

Words you’ll meet (5)

Agent: A model that can call tools, such as search, in a loop and decide when it has enough.
Knowledge graph: Entities (teams, products, people) joined by relations, each traced to the sentence it came from.
Hop: One step along a relation in the graph.
Cross-encoder: A model that reads the question and a chunk together to judge how well they match.
Multimodal: Working with more than plain text: tables, PDFs and images.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 18

Every pipeline so far searched once. Naive RAG is the chapter 3 version: one search, paste the top chunks, answer. Advanced RAG keeps that single pass and tunes it with the chapter 5 levers. To make each type’s strengths easy to see, the retriever here hands over one 16-word chunk.