@agentix-e/embed-code-web

Browser ONNX Runtime adapter for nomic-embed-code — WASM + WebGPU.

Overview

@agentix-e/embed-code-web provides the browser ONNX Runtime inference engine for nomic-embed-code. It uses onnxruntime-web with WebGPU auto-upgrade — falls back to WASM when WebGPU is unavailable.

Installation

npm install @agentix-e/embed-code-web @agentix-e/embed-code-core

Quick Start

import { WebEmbedder } from '@agentix-e/embed-code-web';

// Load tokenizer from JSON (fetch from your server or bundle inline)
const tokenizerJson = await fetch('/models/tokenizer.json').then((r) => r.json());

const embedder = await WebEmbedder.create('/models/nomic-embed-code-v1.5.int8.onnx', tokenizerJson);

const embedding = await embedder.embed('const memo = new Map();');
// → Float32Array(768), L2 normalized

await embedder.dispose();

WebGPU Auto-Upgrade

// WebEmbedder.create() tries WebGPU first, then falls back to WASM:
//   WebGPU available → GPU accelerated (~20ms latency)
//   WebGPU unavailable → WASM backend (~50ms latency)

API Documentation

📚 Full API reference: agentix-e.github.io/embed-code-ts/api/modules/embed-code-web_src_onnx-embedder.html

Key exports:

• WebEmbedder — ONNX Runtime Web embedder implementing IEmbedder with WebGPU auto-upgrade

Performance

| Backend | Hardware | Latency | | ----------------- | -------- | ------- | | onnxruntime-web | WebGPU | ~20ms | | onnxruntime-web | WASM | ~50ms |

License

Apache 2.0