Attention works better on smaller contexts since there's less confounding tokens so even if the LLM can see the entire context, it's better to keep the amount o

Attention works better on smaller contexts since there's less confounding tokens so even if the LLM can see the entire context, it's better to keep the amount of confounding context lower. And at some point the source code will exceed the size of the context window; even the newer ones will millions of tokens of context can't hold the entirety of many large codebases.