RAG gets a language model to answer questions from documents it was never trained on, by finding the right passages and pasting them into the prompt. Most of what makes the searching work predates the models, so this notebook connects it back to ideas that have been around a lot longer.

Two rows. The top runs once whenever the documents change: corpus, chunk, embed, vectors. The bottom runs for every question: question, keyword plus meaning search, fuse, rerank, prompt, answer
The top row runs once whenever the documents change, the bottom row for every question

Read the notebook on GitHub → About a twenty minute read, no account and no API key. Browse the repo.