Graph-and-tree search that grows past your RAM

What it does NGT provides C++ libraries and command-line tools for approximate nearest neighbor search in high-dimensional spaces. It combines a neighborhood graph with a tree-based index, supports adding and removing vectors after initial build, and scales from standard in-memory indices up to billion-object datasets via optional quantization layers. Indices can also be memory-mapped to work with data larger than available RAM.

The interesting bit While most modern vector databases bet everything on graph-only methods, NGT keeps a tree in the mix and offers a clear escalation path from standard indices to heavily quantized ones. Recent releases have added scalar and product quantization alongside half-precision floats, suggesting the project is still evolving from its Yahoo Japan research roots into a practical, multi-modal search backend.

Key highlights

• Hybrid graph-and-tree indexing with dynamic object addition and removal
• Three search tiers: standard NGT, quantized graph (QG), and QBG for billion-scale datasets
• Memory-mapped shared memory support for out-of-core indices
• Wide distance metric support including L1, L2, cosine, angular, Hamming, Jaccard, and inner product
• Bindings for C++, C, Python, Go, Rust, Ruby, and PHP

Caveats

• QG and QBG depend on BLAS and LAPACK, and support fewer distance metrics than the base NGT method
• Shared memory mode lacks locking, so multiple processes must treat the index as read-only
• Contributions require accepting a Contributor License Agreement

Verdict A solid choice if you want a single library that can grow from a few million vectors to billions without abandoning its core architecture. Look elsewhere if you need a fully managed vector database or a zero-dependency, batteries-included Python package.