rtengo

rtengo

Lightweight Go bindings for the RTen WebAssembly runtime. It runs RTen WASM on top of wazero and provides high-level APIs for text embedding, reranking, pruning, and image preprocessing.

Features

• Pure Go WASM runtime via wazero (no CGO)
• RTen WASM embedded by default
• Simple Runtime / Model / Tensor API
• SentencePiece tokenizer support
• High-level task APIs (Embedding / Rerank / Prune)

Requirements

• Go 1.24+
• RTen .rten model file
• SentencePiece tokenizer.model

Install

go get github.com/masacento/rtengo

Quick Start

package main

import (
  "context"
  "log"
  "os"

  rten "github.com/masacento/rtengo"
)

func main() {
  ctx := context.Background()

  // Runtime
  rt, err := rten.NewRuntime(ctx)
  if err != nil {
    log.Fatal(err)
  }
  defer rt.Close(ctx)

  // Model
  modelBytes, err := os.ReadFile("model.rten")
  if err != nil {
    log.Fatal(err)
  }
  model, err := rt.LoadModel(ctx, modelBytes)
  if err != nil {
    log.Fatal(err)
  }
  defer model.Close(ctx)

  // Tokenizer
  tk, err := rten.NewSentencePiece("tokenizer.model")
  if err != nil {
    log.Fatal(err)
  }

  // Embedding
  embedder, err := rten.NewEmbedder(ctx, rt, model, tk)
  if err != nil {
    log.Fatal(err)
  }
  vec, err := embedder.Embed(ctx, "Hello")
  if err != nil {
    log.Fatal(err)
  }
  _ = vec
}

Core API

Runtime / Model / Tensor

rt, _ := rten.NewRuntime(ctx)
model, _ := rt.LoadModel(ctx, modelBytes)
outputs, _ := model.Run(ctx, inputTensor)

shape := outputs[0].Shape()
vec, _ := outputs[0].Float32Data(ctx)

Embedding

embedder, _ := rten.NewEmbedder(ctx, rt, model, tk,
  rten.WithEmbeddingNormalize(true),
  rten.WithEmbeddingPooling(rten.PoolingMean),
)
vec, _ := embedder.Embed(ctx, "text")

Rerank

reranker, _ := rten.NewReranker(ctx, rt, model, tk)
results, _ := reranker.Rerank(ctx, "question", passages)

Prune (OpenProvence, etc.)

pruner, _ := rten.NewPruner(ctx, rt, model, tk, rten.PrunerConfig{Threshold: 0.5})
result, _ := pruner.Prune(ctx, "query", "document")

Image Preprocessing

data, _ := rten.LoadAndPreprocessFile(ctx, "cat.jpg", 224, 224)
inputTensor, _ := rt.NewTensorFloat32([]uint64{1, 3, 224, 224}, data)

Examples

• examples/embedding: text embedding
• examples/rerank: passage reranking
• examples/openprovence: document pruning
• examples/mobilenet: image classification

License

MIT