I built this because I thought it was weird that we treat vector DBs like mutable buckets instead of versioned knowledge bases. RAG apps break easily when embeddings drift, and I wanted `git checkout` for my vectors.

Clamp lets you commit and rollback your vector DB state without copying data. It works by managing metadata flags on points, so "rollbacks" are just instant filter updates, even for large collections.

Basic flow: $ clamp init $ clamp commit docs.txt my_kb "v1 - initial" $ clamp checkout my_kb HEAD~1 # Instant revert

It's a CLI/Python wrapper around Qdrant right now. I'm a student at UW and built this to learn more about vector storage internals.

Code: https://github.com/athaapa/clamp PyPI: https://pypi.org/project/clamp-rag/

Happy to answer questions about the implementation!