Great Question! We have a few options when it comes to how we process a 100s of session replays that each can be a few MBs or greater. The main concept here is that we initially use a grouping system that groups the same signals (such as rage taps or dead taps) that occurred on the same app or website page for many sessions, and admit a random sample of that group into deeper analysis, which is an LLM of choice. We also have other tricks up our sleeves such as reusing session context we already put the effort into processing before among other things. We're working on improving this system though so let us know what suggestions you have!