Build a RAG system

You will have built a RAG system one small function at a time, on a real embedding model and a real chat model: 3 functions that cut documents into chunks, 5 that embed them and find the nearest to a question, 4 that add a keyword ranking and fuse the two, 3 that number the sources and build the prompt, 6 that ask the model, read its citations and abstain when the sources are silent, 3 that check every sentence against the sources, and 2 that put it behind two functions. Each step is a few lines with its own tests, and the earlier functions are provided to the later ones. The models are handed to you: `embedder.embed(texts)` returns one vector per text, `model.chat(messages, system, max_tokens)` returns a reply. Both are real, so the tests for those steps check properties (the right chunk comes first, the fact is in the answer, one call was made) and spend a little of your credits. When every step has passed, the seven files assemble into one system and you can ask it anything about the store, live: it answers with sources, says so when the documents are silent, and flags any sentence it cannot back up.