Course 1 · AI Foundations · Lesson 08
RAG: Open Book Beats Closed Book
The fix promised in Lesson 05. A bare model answers your questions from frozen memory — a closed-book exam. RAG searches real sources first and pastes the best passages onto the desk, so the answer is grounded — and cited.
The one mental model
RAG is four steps: you ask → the system searches real documents (your files, a database, the live web) → pulls the most relevant passages → pastes them onto the desk next to your question → the model answers from what's now in front of it. Search first. Paste onto the desk. Then answer.
Key terms
RAG
Retrieval Augmented Generation — answering after retrieving relevant real text, instead of from memory alone.
Knowledge cutoff
The date the training text was frozen. Anything after it — and anything private — isn't in the model's memory.
Grounding
Anchoring the answer to retrieved passages from sources you chose. The false weight from Lesson 05 loses its grip.
Citation
The answer points at where it came from — page 12, paragraph 3 — so mistakes become checkable. Receipts, not trust.
The misconception to drop
✕
“With RAG / web search the answer is guaranteed correct.”
✓
Open book isn't infallible — it's checkable. The search can grab the wrong passage; the model can misread it. But an answer with a citation carries its receipts, and a wrong one takes thirty seconds to catch.
Put it to work
1
Factual work? Hand it the source — upload the file instead of trusting memory.
2
Ask for citations — and actually click one now and then.
3
Anything after the cutoff → turn web search on.
4
A real pile of documents? Give NotebookLM an afternoon — it answers only from your sources, with a citation on every claim.
Ask the AI Tutor
Pause the video and ask anything from this lesson — the tutor answers from this lesson’s material.
What does RAG actually do, step by step?
Is a cited answer always correct?
When should I use web search vs. plain chat?
What is NotebookLM good for?
Next lesson
09 — Tools, Connectors & Agents