This is insane and obvious at the same time - why is everyone not doing this instead of server side retrieval for webrtc based voice agents?