The elevator pitch for this project is that you give it the URL of a GitHub project, and it starts from a template and iteratively improves on that until it finally succeeds (or fails) to build the package.
🚧 Can only build at most 38 % (14 % verified functionally correct) of all unseen things so far. 🚧
I want to publish a paper based on this work, so if you're interested in that or want to collaborate on related work send me a message.
I could also imagine providing this functionality as a hosted service to customers, so if you are curious about that, please send me an email as well.
To run vibenix with the default terminal UI:
nix develop -c python -m vibenix
You will be asked to pick a model and also provide an API key if required.
Note: There is also a --textual flag for a Textual-based UI, but it is currently unmaintained and not recommended for use.
Models we have tested to perform well are our default model gemini-2.5-flash (38 %),
claude-3-5-haiku-20241022 (32 %) and o3-mini-2025-01-31 (26 %).
Numbers in prarenthesis are raw sucess rate before validation. Manually validated success rate with gemini-2.5-flash is about half, at 14 %.
We would like to target local models in the future, but we are still working on issues to get that working (#42).
Right now vibenix tries to stay within 32k context size, which with 32b class models results in a bit more than 32GB VRAM usage.
