Sweep’s Post

Check our latest blog! Standard vector databases worked great for Sweep when we had a single index, but they became a pain as we started managing multiple indices, frequent updates, and self-hosting. Our backend now consists of a single docker image that handles vector search over hundreds of repositories, each with a few thousand files. We use a simple Redis-based vector database that we run on our own infrastructure, and we'd like to share it with the community. Check out the link in the comments below:

Thank you for Amit Prakash for having me at ThoughtSpot's Generative AI Meetup! Thoughtspot hosts this once a month to share demos, ideas, and tech talks. I’m excited to showcase the newest version of Sweep, an AI developer tool that maintains code for legacy codebases. I’ll also be discussing the technical approaches we’ve used to build the latest version of Sweep (YC S23) (we recently decreased our error rate by xx%). There’s a huge depth in approaches to code search, reasoning over code, and code modification. We’ve focused on extensions to Retrieval Augmented Generation, reliable prompting workflows, and comprehensive benchmarking to all of these. I’m excited to share our learnings and help anyone else working on the same goals. If you’re interested in this, come to the meetup! I’ll post a link in the comments below.

I’m excited to be speaking at the Post-Industrial Institute’s first annual summit in Silicon Valley! This summit will explore how AI is transforming organizations. Our breakout session will cover how AI can boost innovation and accelerate product development. We’ll cover topics like the state of AI programmers, how we measure performance in both business technical metrics, and where Sweep sits in current and future AI workflows. While building Sweep, we’ve been fortunate to partner with large organizations to maintain their legacy codebases. Sweep has been able to match or exceed human developers in certain cases, and we’ve worked with multiple repositories of over 20,000 files. We have space for just one new customer in March. If your engineering team is struggling with bug fixes, security patches, or minor feature work and wants to focus on core product development, reach out!

Sweep is one of the top 10 Show HNs right now! We'd love to hear your feedback/thoughts in the comments. There's a great discussion right now on open models vs OpenAI.

Yesterday we shipped a fully local version of Sweep (YC S23) which uses your PAT to create pull requests from your Github issues! This calls OpenAI directly from your machine without any code entering our servers. We’re excited for more developers to be able to use Sweep without having to get permissions! Here are our docs ⚡: https://lnkd.in/gbU96iFy

Sweep (YC S23) is currently trending on GitHub for Python Repos! Thank you for all of the support, since our release last week we’ve shipped multiple improvements: - Lexical search is now 1.5x faster for all repositories! - We shipped a new modification algorithm! In our testing, this algorithm improves accuracy from 60 → 85%, as well as decreasing latency by 15%! - We launched a new edit button and diff viewer in Sweep’s progress tracker! We’re excited to help onboard new users, please reach out to [email protected] or DM me/Kevin Lu 😀

Sweep (YC S23)'s search over your codebase is now managed through the OpenAI Assistants API! This allows Sweep to: - Read snippets from your codebase if they’re missing. For example, if you ask Sweep to "add docstrings to main.py, then update README.md", Sweep will read README.md even if the search failed to retrieve it! - Remove unnecessary search results! So if Sweep had a bad search result because your test code and main code look similar, Sweep can remove the unnecessary results to improve performance! - Explore additional directories! Now Sweep can see a hidden directory like "src/utils", and look under the "utils" directory to see what’s available! Additional Changes: - Search will be ~2-3x faster using this approach! - Sweep will not support Azure temporarily, we’ll integrate it as soon as the OpenAI API allows it! Overall we’re excited for Sweep to be even better at searching your codebase. If you're interested in using Sweep, join our discord at https://lnkd.in/giDzcY3X!

We just revamped Sweep's refactoring capabilities! Sweep now refactors each function individually and passes in their dependencies. For example, if Sweep is refactoring function f, which calls g in its body, we would pass the function g's body into the context. We would previously include the entire imported file is not very granular and often confuses GPT-4. We also improved our prompts and error handling around GPT-4, and fixed a major bug with file modifications.