TinkerNothing reaches the model: there is no wire path from Docs through the Retriever to the LLM.
Loading the retrieval index…
Chapter 3 of 7
What is RAG?
Pasting everything failed: too big, too stale, no permissions. So the team will hand the model only the few passages each question needs. From this chapter on, the picture on the left is the assistant itself, a pipeline of blocks you build and change.
This chapter: What is the smallest working RAG system, and what does each part do?
You will learn
- The three steps in the name: Retrieve, Augment, Generate
- How a retriever scores every passage against the question
- Why a grounded answer can cite where it came from
Words you’ll meet (4)
- Retriever
- Finds the passages that best match the question and hands the top few to the model.
- Similarity score
- A number for how close in meaning a passage is to the question. Higher is closer.
- Grounded answer
- An answer built from the passages it was given, so it can cite them.
- LLM
- Large language model. It writes text by predicting what comes next, using what it learned in training and what is in its prompt.
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 6
Meet the company assistant. Right now it is just a language model: no documents, nothing wired in. A support lead asks it about refunds.