[REALTIME API] - FEEDBACK - We Built a Star Trek Medical Computer on the Realtime API, It Works 30% of the Time

multitechvisions: > Beyond 10-15 minutes, extraction accuracy drops noticeably. More idle calls, missed data, less precise tool arguments. Medical encounters run 20-40 minutes routinely. This is a critical gap for any real-world clinical deploymen I feel your pain on this, having seen the experience grow worse as time progresses. There is only a 32K input token limit. Have you experimented with calling conversation.item.delete on old items? perhaps with a separate model flagging what can be deleted, and perhaps even introducing a consolidation of several items and deleting them? That is on my TODO list to tackle this problem