@smartpointstech/gnn

Graph Neural Networks on HNSW topology - built in Rust for intelligent graph-based machine learning

Overview

This package provides native Node.js bindings to the spt-gnn Rust crate. It implements graph neural network layers with multi-head attention and message passing on HNSW graph topology, along with adaptive tensor compression and differentiable search. A stub fallback is available when native bindings are not installed.

Installation

npm install @smartpointstech/gnn

Key Features

• SptLayer -- GNN layer with multi-head attention, GRU state updates, and layer normalization
• Differentiable Search -- soft nearest-neighbor lookup returning weighted indices
• Tensor Compression -- adaptive compression levels (half, PQ8, PQ4, binary)
• Continual Learning -- supports EWC and replay buffer patterns
• Hierarchical Forward -- multi-level message passing across HNSW layers

Dependencies

No internal npm dependencies. Uses platform-specific native binaries via optional dependencies (e.g., spt-gnn-darwin-arm64). Built with @napi-rs/cli.

Quick Start

import { SptLayer, differentiableSearch } from '@smartpointstech/gnn';

const layer = new SptLayer(128, 64, 4, 0.1);
const updated = layer.forward(nodeEmbedding, neighborEmbeddings, edgeWeights);

const results = differentiableSearch(query, candidates, 10, 1.0);

API Reference

Exports

Export	Type	Description
SptLayer	class	GNN layer with attention-based message passing
TensorCompress	class	Adaptive tensor compression utility
differentiableSearch	function	Soft nearest-neighbor search with weighted results
hierarchicalForward	function	Multi-level forward pass across HNSW layers
getCompressionLevel	function	Select compression level based on config
init	function	Initialize the GNN runtime
getImplementationType	function	Returns 'native' or 'stub'