Cursor sends the same amount of context regardless if ‘longer context’ is on or off. The answers and solutions to problems are always wrong because (enough) correct context isn’t sent. Even @ing the documents directly doesn’t send the whole thing. Summarizing long documents isn’t possible because it can’t fit more than a couple hundred lines of code.
Who’s idea was it to limit every context on every chat or composer to 10k tokens? That’s barely anything. If you’re concerned people will rack up big bills with it or not know what they’re doing, hide it behind an advanced setting. Or force users to use their own OpenAI Endpoint or Google API key.
There’s 0 reason why my Flash 2.0 should be getting the most basic questions wrong in Cursor when just copying and pasting my file into AIStudio gets it right first time, all because Cursor limits requests to 1/100th of the possible model context window for absolutely no reason and thinks I’m too stupid to set a new limit myself.