git-vdb

A Git-native embedded vector database.

git-vdb stores vectors, payloads, metadata, and search indexes as immutable Git objects. Equivalent data produces the same root tree, while ordinary Git commits and refs provide optional naming, history, transport, and atomic writes.

Install

cargo add git-vdb       # Rust library
cargo install git-vdb   # CLI

The project also provides a pinned Nix development and build environment:

nix run github:nishu-builder/git-vdb -- --help

Rust quick start

use git_vdb::{CollectionConfig, Point, Query, SnapshotEngine};

# fn main() -> git_vdb::Result<()> {
let engine = SnapshotEngine::ephemeral()?;
let snapshot = engine.build(
    CollectionConfig::new(2),
    vec![
        Point::new("east", [1.0, 0.0]),
        Point::new("north", [0.0, 1.0]),
    ],
)?;

let result = snapshot.query(Query::exact([0.9, 0.1], 1))?;
assert_eq!(result.points[0].id.to_string(), "east");
# Ok(())
# }

See the API documentation for named collections, immutable snapshots, filtering, approximate search, validation, history, and materialization.

CLI quick start

git-vdb init ./vectors.git --bare
git-vdb --repo ./vectors.git collection create products --dimension 3
git-vdb --repo ./vectors.git upsert products --input points.jsonl
git-vdb --repo ./vectors.git query products \
  --vector query.json --limit 10 --exact --with-payload

Inputs and outputs are JSON or JSON Lines. Run git-vdb --help or git-vdb <command> --help for the complete command reference.

Model

tree   = deterministic database state
commit = history describing a state transition
ref    = mutable name for the current state

Use SnapshotEngine when another system owns naming and persistence. Use Database, Collection, or the CLI when you want named collections backed by commits and compare-and-swap refs. Both layers produce the same canonical roots.

Core capabilities include:

typed string and unsigned-integer IDs, dense f32 vectors, and JSON payloads;
exact cosine search and deterministic IVF-flat approximate search;
Boolean payload filters, retrieval, counts, upserts, and deletion;
immutable historical reads, diffs, validation, and portable snapshots;
bare or non-bare Git repositories with no server or hidden external state.

Approximate search is explicitly bounded and may omit globally better points; query results report the selected mode and work statistics.

Documentation

Persisted format versions and the crate's Semantic Version are separate compatibility boundaries. New roots use version 2; version-1 roots remain readable and mutable.

Development

nix develop
nix flake check

Licensed under the Apache License 2.0.