Hi maybe I’m wrong but It seems that cursor won’t really submit the entire codebase of my project resulting in incomplete response. It is a bug or “a feature” 
?
I’m on most recent update on fedora 41 flatpak.
Unfortunately, the underlying AI models Cursor users are not yet powerful enough for us to submit an entire codebase, line for line, and have the model perform intelligent and informed edits on it.
While some models may technically be capable of having a whole codebase sent to them (like Gemini 1.5 Pro, which accepts up to 2 million tokens), the actual effective input, where the model has full knowledge of the entire input, is much smaller.
As such, when you use @codebase, Cursor uses a semantic ordering algorithm to rank each file in your codebase based on how relevant it may be to what you are asking the model to do.
In this case, out of 61 searched files, even the highest-scoring file only has a 0.35 score, as your prompt was not asking about anything specific in your codebase.
Once we rank the files, as you can see in the “Final Codebase Context” section, we then only send the most relevant chunks of code to the AI model you’ve chosen for it to generate a reply.
While this is the best solution for codebase-wide edits, our ‘Agent’ feature can do so in a much smarter way, as it has the ability to traverse and investigate your codebase all on its own, much like a human would. I’d highly recommend trying the agent if you haven’t already, as it far outdoes the @codebase feature for adding new features or editing existing code.
1 Like
Agree. the new “no composer” makes more hard to understand if the agent knows all the code. if crucial when you are working in HUGE projects alone or in a refactoring. i made a LOT of documentation of the project and of the codebase and was ignored repetidly or randomly. the status of de codebase HAS TO BE clear