@vivantel/virage-embedder-transformers

Local, offline embedding provider for @vivantel/virage-core using @huggingface/transformers. No API key required — models run entirely on your machine.

Installation

npm install @vivantel/virage-embedder-transformers

Usage

import { TransformersEmbedder } from "@vivantel/virage-embedder-transformers";

const embedder = new TransformersEmbedder({
  model: "Xenova/all-MiniLM-L6-v2",
  dimensions: 384,
});

The pipeline initializes lazily — the model is downloaded on the first embed() call and cached locally.

JSON config

{
  "embedder": {
    "package": "@vivantel/virage-embedder-transformers",
    "config": {
      "model": "Xenova/all-MiniLM-L6-v2",
      "dimensions": 384
    }
  }
}

Recommended models

| Model | Dimensions | Notes | | ---------------------------------------------- | ---------- | ------------------------- | | Xenova/all-MiniLM-L6-v2 | 384 | Fast, good quality, 80 MB | | Xenova/all-mpnet-base-v2 | 768 | Higher quality, 420 MB | | Xenova/paraphrase-multilingual-MiniLM-L12-v2 | 384 | Multilingual |

Any ONNX-compatible model on the HuggingFace hub can be used.

Options

| Option | Type | Description | | ------------ | ------------------- | ----------------------------------- | | model | string | HuggingFace model ID (required) | | dimensions | number | Output dimensions (default: 384) | | device | "cpu" \| "webgpu" | Inference device (default: "cpu") | | cacheDir | string | Local model cache directory |

First run

On the first embed() call the model weights are downloaded from HuggingFace and cached locally (~80–420 MB depending on the model). Subsequent runs use the cached model.

License

MIT