RAG Explained — Why Some AIs Can Quote Your Notes

RAG (Retrieval-Augmented Generation) lets AI work with documents it didn't train on. Most school AI tools use it.

What to actually do

• Step 1: chop the document into chunks
• Step 2: turn each chunk into an embedding
• Step 3: when you ask a question, fetch the most relevant chunks and feed them in

The big idea: RAG lets AI work with stuff it never saw before. Retrieve the right chunks, then write the answer.

RAG (Retrieval-Augmented Generation) is when AI searches the web first, then summarizes results. That's why Perplexity feels current and ChatGPT sometimes feels dated.

Some examples

• Perplexity uses RAG — it searches before answering.
• Plain ChatGPT only knows what it was trained on.
• RAG can still summarize search results inaccurately.
• Always click sources even when RAG cites them.

Try it!

Ask Perplexity 'what happened in the news yesterday?' then ask ChatGPT the same. Notice how different they feel.

RAG (Retrieval-Augmented Generation) is the technique behind almost every serious AI product: NotebookLM, Custom GPTs with files, ChatGPT Search, Perplexity, almost every company chatbot. The idea: when you ask a question, the system FIRST searches a database of relevant documents, THEN feeds the top hits to the LLM as context, THEN the LLM answers using only those documents. This solves the hallucination problem (the AI is grounded in real text) and lets companies build AI on their own private data. Every job listing for 'AI engineer' mentions RAG.

Some examples

• NotebookLM is a textbook RAG product — your uploaded PDFs become the retrieval source; the AI cannot answer outside them.
• ChatGPT Search is RAG at internet scale — Bing retrieves pages, GPT generates the answer using them.
• Vector databases (Pinecone, Weaviate, ChromaDB) store the documents in a way that enables fast semantic search — that's the 'how.'
• When a company says 'we built ChatGPT for our help docs,' they mean RAG — usually a weekend project for one engineer.

Try it!

Build your own mini-RAG: in ChatGPT or Claude, upload 3 PDFs of class notes, then ask only questions about them. Notice the answers come straight from the PDFs and refuse to wander. You just used RAG.