Stupid question that I could look up myself but asking here…

Is context represented in by the original text, or translated into token indexes? Because if it’s in the original form, it seems obvious that either the server should be compressing into token indexes, or on handshake the server gives the client its dictionary and on request it translates it