Explain how RAG works end to end.

Indexing (offline): split the documents into chunks, turn each chunk into an embedding, and store them with their text and metadata in a vector index.
Query: embed the user's question with the same model and fetch the nearest chunks (often mixed with keyword search).
Optional steps: filter by permissions, rerank the candidates, trim to the few best.
Generation: put the chunks and the question in the prompt, tell the model to answer only from them and to cite them, and to say when the answer is not there.
Updating knowledge means re-indexing the changed documents, not retraining the model.
Evaluate retrieval (is the right chunk found?) and generation (does the answer follow the chunks?) separately.