Lesson 0.1

What RAG Actually Is

A language model that answers from sources you hand it, instead of from memory. That one change is the whole idea.

concepts 8 min read

A closed-book exam, made open-book.

Ask a language model a question and, by default, it answers from memory: whatever it picked up while training, squeezed into its weights. That works for common knowledge and fails the moment you ask about something specific, recent, or private. It does not say "I do not know." It writes a fluent, confident answer that may be wrong.

Retrieval-augmented generation (RAG) turns that closed-book exam into an open-book one. Before the model answers, you go and find the passages that actually address the question, and you put them in front of the model. Same model, but now it reads from a source instead of reaching into memory.

Drag the document into the model

the model's context LLM the same model launch-notes.md

Q  What shipped in the Q3 launch?

A  "I'm not certain, possibly the mobile app."

guessing from memory

The model did not get smarter. It got a source to read. RAG is the work of finding the right source and handing it over before the model speaks.

Retrieve, then generate.

The name says it in order. First retrieve: find the passages relevant to the question. Then generate: let the model write the answer using those passages as its context. The whole system is the plumbing that makes those two steps fast and reliable.

question → find the closest passages → paste them into the prompt → grounded answer + citation

Four hops, every question. Lesson 0.3 walks each one; the rest of the course builds them.

What RAG is not

Re-training the model

No weights change. The model is untouched; only its input changes. That is why RAG can use fresh or private data the model never saw.

Just search

Search returns documents and stops. RAG goes one step further: it reads them and writes an answer, with a pointer back to the source.

A guarantee of truth

Grounding makes the model far more reliable, not infallible. It can still misread a passage. Unit 4 looks at when it slips.

Key takeaways

1

RAG makes a model answer from sources you provide, not from its own memory.

2

It is two steps in order: retrieve the right passages, then generate the answer from them.

3

Nothing about the model is re-trained. The leverage is entirely in what you put in front of it.