VaneDB

VaneDB

Embeddable vector database for edge AI

---

Header-only C++20 vector database with SIMD acceleration. Runs on Linux, macOS, Windows, iOS, and Android.

‍Two implementations. VaneDB is maintained as both C++ and Rust under the @vanedb org. This repo is the C++ header-only version — drop a header into any CMake project, no Rust toolchain needed. For the Rust crate (with Python/PyO3 and WASM bindings), see vanedb/vanedb.

Why VaneDB?

Feature	VaneDB	FAISS	hnswlib	Pinecone
Header-only	Yes	No	No	N/A
Mobile/Edge	Native	No	Partial	No
Dependencies	Zero	Many	Few	Cloud
Binary size	<100KB	200MB+	~1MB	N/A
GPU (Metal)	Yes	No	No	N/A

Perfect for: Mobile AI apps, Obsidian/Logseq plugins, edge devices, offline-first applications.

Features

• SIMD-optimized: ARM NEON, x86 AVX2 (~100ns for 768d vectors)
• Multiple indexes: Brute-force, HNSW, Memory-mapped
• GPU acceleration: Metal (Apple Silicon). CUDA (NVIDIA) is experimental — kernel source only, not yet wired into the build
• Thread-safe: Concurrent reads with std::shared_mutex
• Python bindings: NumPy integration, GIL-safe

Quick Start

#include "core/vector_store.h"
 
vanedb::VectorStore store(768, vanedb::DistanceMetric::COSINE);
store.add(1, embedding);
auto results = store.search(query, 5);  // top-5 nearest neighbors

#include "core/hnsw_index.h"
 
vanedb::HNSWIndex index(768, vanedb::DistanceMetric::COSINE, 100000);
index.add(1, embedding);
auto results = index.search(query, 5);
index.save("index.bin");

import vanedb_py as vanedb
import numpy as np
 
index = vanedb.HNSWIndex(768, vanedb.DistanceMetric.COSINE)
index.add(1, np.random.rand(768).astype(np.float32))
ids, distances = index.search(query, 10)

Build

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
ctest --test-dir build --output-on-failure

Documentation

• Full API Guide - Detailed usage, Python bindings, mobile builds
• CHANGELOG - Version history

License

MIT License - see [LICENSE](LICENSE) for details.