This guide will help you set up pre-downloaded ONNX models for the fastest embedding generation with zero network overhead.

# Navigate to the models directory
cd search/models

# Run the download script
./download-model.sh

# Wait for download to complete (may take a few minutes depending on connection)
# The model is ~90MB total

If the script doesn't work, you can download manually:

cd search/models

# Clone the Hugging Face repository
git clone https://huggingface.co/Xenova/all-MiniLM-L6-v2

# Remove .git directory to save space
rm -rf all-MiniLM-L6-v2/.git

Or download individual files from: https://huggingface.co/Xenova/all-MiniLM-L6-v2/tree/main

Required files:

• onnx/model.onnx (~23MB) - Main ONNX model
• onnx/model_quantized.onnx (~6MB) - Quantized version (optional, faster)
• tokenizer.json - Tokenizer vocabulary
• tokenizer_config.json - Tokenizer configuration
• config.json - Model configuration

Check that the model files exist:

ls -lh all-MiniLM-L6-v2/
ls -lh all-MiniLM-L6-v2/onnx/

You should see:

all-MiniLM-L6-v2/
├── onnx/
│   ├── model.onnx              (~23MB)
│   └── model_quantized.onnx    (~6MB)
├── config.json
├── tokenizer.json
├── tokenizer_config.json
└── ...

In search/index.ts, make sure this line is active:

import { searchStrategy, generateEmbeddings } from "./transformers-local-onnx.ts";

Comment out other strategies:

// import { searchStrategy, generateEmbeddings } from "./transformers-cosine.ts";

Edit search/transformers-local-onnx.ts to customize:

// Use quantized model (faster, slightly less accurate)
const USE_QUANTIZED = true;  // Default: false

// Model path (change if you want to use a different model)
const MODEL_PATH = new URL("./models/all-MiniLM-L6-v2", import.meta.url).pathname;

The local ONNX version:

• ✅ No network calls - works completely offline
• ✅ No downloads on first run - instant startup
• ✅ No isolate loading delays
• ✅ Same accuracy as the cached version
• ✅ Perfect for serverless/edge deployments

• Upload the model files to Val.town's blob storage or use a CDN
• Update MODEL_PATH to point to the remote location
• Or use the regular transformers-cosine.ts which downloads to isolate cache

• Include the models in your deployment package
• Models will be read from the file system
• Works perfectly with the local strategy

• Workers have a 1MB size limit, so the local strategy won't work
• Use cloudflare-bge-cosine.ts instead (uses Cloudflare AI)

• Include the models directory in your Docker image
• Works perfectly with local file system access
• Fastest option for these platforms

• Ensure model files are downloaded: ls -lh search/models/all-MiniLM-L6-v2/
• Run the download script: ./search/models/download-model.sh
• Check file permissions: chmod -R 755 search/models/all-MiniLM-L6-v2/

• Make sure you're using the correct import in search/index.ts
• Check that the file path is correct

• Try the quantized model: set USE_QUANTIZED = true in transformers-local-onnx.ts
• Check if model files are on a fast disk (SSD recommended)

• The model requires ~100-200MB RAM when loaded
• Use the quantized version to reduce memory usage

You can use other ONNX models from Hugging Face. Popular options:

• all-MiniLM-L6-v2 (default) - 384 dims, 23MB, fast
• all-mpnet-base-v2 - 768 dims, 420MB, more accurate
• paraphrase-multilingual-MiniLM-L12-v2 - 384 dims, 470MB, multilingual

To use a different model:

1. Download it to search/models/[model-name]
2. Update MODEL_PATH in transformers-local-onnx.ts
3. Update embedding dimension in your data pipeline if needed

To remove the model and reclaim disk space:

cd search/models
rm -rf all-MiniLM-L6-v2

Then switch back to a different strategy in search/index.ts.