"Can AI answer questions from our documents?" is the question every company asks. The answer is RAG — retrieval-augmented generation — and this lab makes it visible: watch a bot search a real document library, pick its chunks, fill its context window and write an answer. Then watch it fail — hallucinate, truncate, go stale — and learn to fix every failure. Six levels, then a sandbox.
Copilots, enterprise search, support bots trained "on your knowledge base" — under the hood they are all the same five-step machine. Learn it here, where every step is visible, and no vendor demo will ever fool you again.
Documents get sliced into pieces before anything else happens. Slice too big and the model drowns in irrelevant text; too small and the meaning gets cut in half. Level 3 lets you feel both failures.
Your question is compared against every chunk and the closest matches win. It's search, not magic — and when the right chunk isn't retrieved, the model literally cannot know the answer.
The model only reads what retrieval hands it — a limited budget of text. Every irrelevant chunk in the window is budget stolen from the relevant one.
The single most important switch in RAG: may the model answer only from the documents, or improvise when they're silent? Level 5 shows you what confident improvisation looks like.
Answers that point to their sources can be checked; answers that don't must be trusted. In a business, "trust me" is not a feature. Citations turn the bot from an oracle into a librarian.
Everything you just learned — chunking, retrieval, grounding, citations — is exactly what we tune when we build document-answering assistants for real teams. Cocoon trains your people to use them well, and builds them properly.