Comment by immibis2

Something like: An LLM can encode a query into a vector to match against a vector database of indexed pages.

You have to actually engineer that though, you can't just throw random stuff into ChatGPT and expect good results.