Retrievalthe open-book exam (RAG)
Embed the question, nearest-neighbor a document space, paste what you find into context.
Before this
Next in Frontier AI
AgentsFrontier AI, stop 9 of 102 minThis opens up
Nothing yet. No lesson builds on this one so far.
Transcript
It follows the voice. Click a sentence to jump there.
Ask about yesterday's earnings call and you hit problem one: everything the model knows is baked into its weights. Those weights froze the day training ended. Yesterday doesn't exist in there.
Problem two is worse: the model has no idea what it doesn't know. Revenue jumped, the CEO stepped down, margins collapsed — all invented, all delivered with identical confidence. Nothing in the machinery flags a guess as a guess.
The fix starts with embedding geometry. Every passage in your documents already lives as a point in meaning-space. Embed the question the same way and it lands on the same map. The passages nearest to it are the ones that matter.
Step two: paste those retrieved passages into the prompt, directly ahead of the question. No retraining, no new weights — just fresh text in the context window. The exam is now open-book.
Attention does the rest, the same attention that was already there. Answer tokens reach straight into the retrieved passages, so every claim is anchored to a real source. The citations aren't decoration — the model read those lines.
Retrieval-augmented generation: a search engine and a Transformer as one system. Yesterday's earnings call was never in the frozen weights — and it never had to be. When a model needs knowledge it wasn't trained on, you don't grow it a bigger brain. You hand it a library card.