What if instead of pasting your entire codebase into a chat window, you let the AI search for what it actually needs?
That's the premise behind git-semantic, an open-source experiment proposing a replacement for Claude Code's default context-loading approach. The project is early-stage - a small repository published by developer ccherrad - but it addresses a real friction point in AI-assisted coding.
Claude Code currently works by pulling code files, function definitions, and related context directly into its conversation window (the total amount of text the AI can process at once, roughly 200,000 words in Claude's case). On large codebases, this fills fast. You end up with a window stuffed with tangentially relevant code, leaving less capacity for the actual problem being solved.
git-semantic uses embeddings - numerical representations of what code means and does, rather than just what it literally says - to find the most relevant snippets before passing anything to Claude Code. Instead of including everything and hoping the model focuses on the right parts, the tool attempts to surface only what matters.
The project has no published benchmarks and the implementation is minimal. There's no direct comparison showing measurably better code suggestions. This is genuinely experimental work.
The underlying technique - called RAG, or retrieval-augmented generation - is how enterprise search tools handle large document collections. You don't load all 10,000 files into the AI; you search for the relevant 20 first. Applying that pattern to local code editing is a natural direction to explore.
The project isn't production-ready. But it raises a question worth sitting with: is more context always better, or does better-selected context outperform more of it?