More likely some clever take on RAG. There’s no way that 1M context is all available at all times. More likely parts of it are retrievable on demand. Hence the retrieval-like use cases you see in the demos. The goal is to find a thing, not to find patterns at a distance