Spend time reviewing code generated by AI, not writing it.
Sweep allows you to create and review GitHub issues with easy. Simply describe any issue and Sweep will do the rest. It will plan out what needs to be done, what changes to make, and write the changes to a PR.
For the best experience, install Sweep to one of your repos and see the magic happen.
sweep.mp4
Setting up Sweep is as simple as adding the GitHub bot to a repo, then creating an issue for the bot to address.
- Add the Sweep GitHub app to desired repos
- Create new issue in repo, like "Sweep: Write tests"
- "👀" means it is taking a look, and it will generate the desired code
- "🚀" means the bot has finished its job and created a PR
For more detailed docs, see 🚀 Quickstart.
We were frustrated by small tickets such as simple bug fixes, annoying refactors and small features that could just be fed to ChatGPT. So we fed it to ChatGPT.
Unlike copilot, this can solve entire tickets and can be parallelized: someone can spin up 10 tickets and have it solve them all at once.
- GPT-4 32k 0613 (default) / Claude v1.3 100k
- ActiveLoop DeepLake for Vector DB with MiniLM L12 as our embeddings model
- Modal Labs for infra + deployment
- Vector search using DeepLake
- CoT internal search using GPT Functions
- Issue comment reply handling
- PR auto self-review + comment handling (effectively Reflexion https://arxiv.org/abs/2303.11366)
We're currently working on responding to linters and external search. For more, see 🗺️ Roadmap.
Thank you for your contribution!
and of course, Sweep!