Question on repository level code completion

[ZuoyunZheng]

Hi,
newbie here. With my usage of this (very useful and fun) plugin, i always find myself having to open up the relevant files in split for the model to be able to complete the code as i wish.

I wish there was some way to also pass on all the code in my repository to the model so that it could perform some sort of RAG. My question is, how realistic is it that this gets implemented in the plugin? If not, how would you implement something like this?

Thanks in advance!

[ggerganov]

For the short term I don't think we will add repository-level RAG. It requires a complicated orchestration logic, that would be a lot of effort to maintain and implement correctly.

There are some simpler approaches to explore that would improve the "repo awareness" of the completions and would not require a sophisticated implementation. For example, a better chunk selection logic for the extra context might improve the results significantly.

[bhupesh-sf]

agreed, though repo level RAG would be best but that's quite a moving target to implement and maintain. But having better chunk selection logic which can cover whatever max can be covered for the repo will be great