Without RAG
Made up“Yes. Enterprise customers can get a full refund within 60 days of purchase, no questions asked.”
No documents reached the model, so it answered from general knowledge and made up a policy that sounds plausible.
Module 1 · AI
How AI assistants answer from your company’s documents
Retrieval-Augmented Generation (RAG) is how most workplace AI assistants find the right policy, contract or price before they answer. You’ll build one block by block, watch it fail, and fix it.
- 7 chapters
- No code
- Maths optional
- Runs in your browser
The whole idea in one question
Same model. Same question. Different answers.
An employee asks “Can enterprise customers get a refund?”
With RAG
Correct“No. The refund policy does not apply to enterprise contracts. Enterprise customers who want money back raise a billing dispute with their account manager, who reviews it within ten business days.”
From refund-policy-2025The refund policy does NOT apply to enterprise contracts.
These are the course’s own recorded answers. The only thing that changed is what reached the model.
How it works
Three steps hidden in the name
1 Retrieve
Search every document for the passages closest in meaning to the question. Keep the best few.
2 Augment
Put those passages into the prompt, next to the question, so the model reads them first.
3 Generate
The model writes the answer from what it was handed, and can point to where it came from.
The story
You’ve just joined Larkspur Supply Co.
Larkspur sells office and warehouse hardware, from scanners to steel shelving. Its staff lose hours hunting through policies, contracts, price lists and meeting notes, so the company wants an AI assistant that answers from those documents. You’re on the team building it.
- ~4,000
- business customers
- ~260
- staff
- 47
- documents on the drive
- 2
- warehouses
Questions waiting for the assistant
“What is the hotel nightly limit for business travel?”
A retired policy still sits in the drive · chapter 5
“it's broken again, who do I even call?”
Nothing in the documents is worded like this · chapter 5
“How long does Larkspur keep customer payment card data?”
A look-alike note uses the same words · chapter 5
“Who approves purchases for the team that owns the Kestrel K2 scanner integration?”
The answer needs two documents joined · chapter 6
“How much does a Kestrel K2 scanner cost per unit when ordering 50 or more?”
The answer is inside a table · chapter 6
Your path
Seven chapters, one assistant
How RAG came to be · done
Why can’t an AI model simply answer questions about your company? And how did the industry arrive at RAG between 2020 and 2026?
Why not stuff it all in the prompt? · done
Why does “just give it everything” break down, even with today’s huge context windows?
What is RAG? · done
What is the smallest working RAG system, and what does each part do?
AddsDocsRetrieverLLM
The layers of a RAG system · done
What happens between “a pile of documents” and “the passages the model reads”, and where can it break?
AddsChunkerEmbedderIndexContext assembler
Making RAG accurate · done
When a RAG answer is wrong, how do you tell whether retrieval or the model failed, and which lever fixes it?
AddsQuery rewriteRerankerEvaluator
Types of RAG · done
Which kind of RAG fits which kind of problem, and what does each one cost?
AddsAgentGraph
Sandbox · done
Can you build, break and tune a RAG pipeline on your own, and pick the right type for a real team’s problem?
Who it’s for
Product managers, engineers, analysts, and anyone whose company is rolling out AI assistants who wants to know how they really work, where they fail, and what the fixes cost. No background needed.
How a chapter page works
The picture on the left is live. The story on the right asks you to do one thing in it at a time. Hints appear when you’re stuck, and puzzles have a “Show me”. Progress stays in this browser.