Embeddings and RAG, without the jargon

Two terms come up constantly once you go past basic chat: embeddings and RAG. Together they’re how you get an assistant to answer reliably from your material — a policy library, a product manual — rather than from its frozen training.

Embeddings: meaning as coordinates

An embedding turns a chunk of text into a long list of numbers that captures its meaning. Texts about similar things land near each other in that number-space, even when they share no words. That’s what makes “find me passages about refund deadlines” work when the document actually says “the window for returns”.

RAG: retrieve, then generate

Retrieval-Augmented Generation is the pattern built on top. When you ask a question, the system first retrieves the most relevant chunks from your documents (using embeddings to find them), then hands those chunks to the model and asks it to generate an answer grounded in them.

The payoff is big: answers cite your real content, the model is far less likely to invent things, and you can update knowledge by changing documents instead of retraining anything.

RAG doesn’t make a model smarter — it makes it better informed. Quality depends on what you retrieve, so messy source documents produce messy answers.

It’s also why context windows matter here: retrieved chunks have to fit alongside the question and the answer.