Comment by vermorel

With 1M tokens, if snapshotting the LLM state is cheap, it would beat out-of-the-box nearly all RAG setups, except the ones dealing with large datasets. 1M tokens is a lot of docs.