GitDB – GPU-accelerated vector database with Git-style version control

Name: GitDB – GPU-accelerated vector database with Git-style version control
Availability: InStock
Author: vkaufmann

by vkaufmann·Mar 14, 2026·3 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainBold BetShip It

Git-style version control for vectors when Chroma and Qdrant can't time-travel.

Strengths

•Native git log, diff, branch, merge operations on vector embeddings — genuinely novel approach.
•CEPH CRUSH placement enables coordinatorless horizontal scaling without central server.
•Universal ingest swallows SQLite, MongoDB, Parquet, PDF with single command.

Weaknesses

•Zero stars, eight commits — embryonic compared to established vector databases.
•GPU acceleration claimed but no benchmarks showing actual performance gains.

Post Description

GitDB is a version-controlled vector database. Branch, merge, diff, and time-travel your embeddings the same way git handles source code. No server, no Docker — pip install gitdb-vectors and go.

What makes it different:

- Version control is native — git log, git diff, git branch, git merge for vectors. Roll back to any commit. Cherry-pick by meaning. - Time-travel queries — db.query_text("revenue", at="v1.0") searches an old snapshot. - CEPH CRUSH placement — deterministic data routing. Scales horizontally by adding peers, no coordinator. - P2P distributed — peers sync over SSH like git remotes. Each node is a full shard. - FoundationDB features — hooks, transactions, watches, secondary indexes, schema enforcement. - Universal ingest — swallow entire SQLite databases, MongoDB exports, CSV, Parquet, PDF with one command: db.ingest("legacy.db") - CLI-first — works from terminal like git: gitdb init && gitdb add --text "doc" && gitdb commit -m "init" - Embedded — no server process. Import and query in 3 lines of Python.

21 modules, 13,150 lines, 394 tests. GPU-accelerated (MPS/CUDA/CPU).

GitHub: https://github.com/VincentKaufmann/GitDB

PyPI: pip install gitdb-vectors