LLMs are great at historical data. They however still suck at being kept up to date. I would expect that someone (if not already has) will find a cost efficient way to continuously index new data, but it would be interesting to see how it stacks up with the computing power needed for modern search index pipelines.