milestone... making llm aware of entire code base

[theesfeld]

Perhaps by using the git repository information, it would be great for the llm to have awareness or knowledge of the entire code base, to know about custom imports, functions in other files, utility code files, etc.

I know this is far fetched, so not really an issue but more of a dream :)

[yetone]

‌We can refer to the implementation plan of Aider's repo map.

[pidgeon777]

It would be a big feature, for sure!

[theesfeld]

it would be game changing. this is the feature no one has but everyone wants

[aarnphm]

my guess is that this would only works with claude or models that support 1M+ tokens

[mrloop]

Here's the blog post for the repository map for an overview. I use a few different LLM as auto complete, but don't think any of them take the approach mentioned in the blog, instead sending the open buffer(s) for context. With the repo map you'd specify number of token and the repo map would include most popular function definitions in it, this way you support different models with different context windows. I've started building a lua module to do this, the intention was to have a single module that could be used by other projects, mostly thinking of the ones I use for auto complete. Would love it if avante built it, then it could be extracted to standalone module. Or I if get round to building mine it could be used here, and in other places.

[theesfeld]

just wanted to throw this out there https://github.com/Doriandarko/claude-engineer

[mrloop]

Started work on repo-map.nvim

[aarnphm]

from #456 too track open buffer as an option.

[SynapticSage]

bump! also very interested in this feature.

I do love the idea of defaulting to open buffers. though would be nice to have drop and add file commands like aider.