Installation

Install the package, set up the native TensorFlow backend, and verify everything works.

Install

Deno (recommended)

deno add jsr:@casys/shgat

npm

npx jsr add @casys/shgat

Published on JSR, the modern JavaScript registry.

Requirements

Dependency	Minimum Version	Notes
Deno	2.0+	FFI requires --unstable-ffi flag
libtensorflow	2.15.x	Native C library (see below)
Linux x86_64	Ubuntu 20.04+ / Debian 11+	macOS x86_64 also supported
CUDA (optional)	11.8+	Only for GPU acceleration

TensorFlow Installation

@casys/shgat uses the C TensorFlow library directly via Deno.dlopen. You need libtensorflow installed on your system.

Option A: Install script

./scripts/install-libtensorflow.sh        # CPU (default)
./scripts/install-libtensorflow.sh gpu     # GPU (requires CUDA 11.8+)

Option B: Manual install

Linux

# Download libtensorflow 2.15.0 (CPU)
curl -L https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-2.15.0.tar.gz \
  -o /tmp/libtensorflow.tar.gz

# Extract to /usr/local
sudo tar -C /usr/local -xzf /tmp/libtensorflow.tar.gz

# Update library cache
sudo ldconfig

# Verify
ls /usr/local/lib/libtensorflow.so

macOS

# Download libtensorflow 2.15.0 (CPU)
curl -L https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-darwin-x86_64-2.15.0.tar.gz \
  -o /tmp/libtensorflow.tar.gz

# Extract to /usr/local
sudo tar -C /usr/local -xzf /tmp/libtensorflow.tar.gz

# Verify
ls /usr/local/lib/libtensorflow.dylib

Verify it works

Create a test-shgat.ts file:

import { createSHGAT, type Node } from "@casys/shgat";

const nodes: Node[] = [
  { id: "tool-a", embedding: new Array(1024).fill(0.1), children: [], level: 0 },
  { id: "tool-b", embedding: new Array(1024).fill(0.2), children: [], level: 0 },
  { id: "group",  embedding: new Array(1024).fill(0.15), children: ["tool-a", "tool-b"], level: 1 },
];

const model = createSHGAT(nodes);
console.log(`Nodes registered: ${model.getToolCount()} leaves, ${model.getCapabilityCount()} composites`);

const scores = model.scoreNodes(new Array(1024).fill(0.1));
console.log("Top result:", scores[0]?.nodeId, scores[0]?.score.toFixed(4));

model.dispose();

Run it:

deno run --unstable-ffi --allow-ffi test-shgat.ts

Tip

If you see the node count and a score, the native TensorFlow backend is working. Head to the Quick Start to score real nodes.

Note

SHGAT uses libtensorflow via Deno FFI (Deno.dlopen). No WASM, no npm tf packages, native C performance. The --unstable-ffi and --allow-ffi flags are required for Deno to load the shared library.

Next up

Score your first real nodes and see ranked results: Quick Start.