GitHub is building a search tool that reads your whole codebase

GitHub Next has quietly published a paper on Knowledge Compressor, a model that digests an entire repository and turns it into a searchable embedding. The idea is simple: instead of writing prompts that say "look at files X, Y, and Z," you point the model at the repo and it pulls out what it needs. It's not a chatbot. It's a retrieval layer sitting on top of your code.

The tooling is built on Vercel's new run platform, which lets you execute models on demand without managing GPUs. You spin up a container, pass it the repo, and the model writes back its findings. It's the kind of infrastructure work that usually happens inside one company before anyone else notices.

The bigger picture is that everyone is trying to solve the same problem: how do you make an AI actually read your codebase without you writing fifty lines of prompt engineering? The answer so far is embeddings plus search. Knowledge Compressor is just one take on that, but it's the kind of thing that will show up in Copilot, Cursor, and whatever comes next.

Why this matters for us: if your shop is still paying for an engineer to read through code to answer a simple question, this tool is one step closer to making that a thing of the past — and the people who get it first will be the ones asking the hardest questions about their repos.