Get the fastest semantic search running in 3 steps.

cd search/models
./download-model.sh

Time: ~2-5 minutes (depending on internet speed)
Size: ~90MB download
What it does: Downloads all-MiniLM-L6-v2 ONNX model from Hugging Face

Edit search/index.ts:

// Comment out the current strategy
// import { searchStrategy, generateEmbeddings } from "./transformers-cosine.ts";

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

cd models
deno run --allow-read --allow-env --allow-net --allow-ffi test-local-model.ts

Note: --allow-ffi is required for ONNX runtime native bindings.

Expected output:

📂 Loading local ONNX model from: /path/to/models/all-MiniLM-L6-v2
   Using full model
✅ Local ONNX model loaded successfully:
   npm import: 102ms
   pipeline load: 137ms (from local files)
   total model load: 239ms

✅ Generated embeddings successfully!
   Dimensions: 384
   First 5 values: [-0.0457, -0.0109, -0.0935, ...]
   Time: 247ms

Test 2: Generate embeddings for multiple queries (testing cache)
✅ "How to use Groq API?"
   Time: 3.87ms (cached pipeline)

Note: You may see a harmless error at the very end (mutex lock failed) - this is a known ONNX runtime cleanup issue in Deno that only affects standalone scripts that exit immediately. Your long-running server (main.tsx) won't have this issue. All functionality works correctly!

Your search is now running with:

• ✅ No network calls (100% offline)
• ✅ No API keys needed
• ✅ ~10-30ms query time
• ✅ ~60-80ms total search time

Start your app and check /search/test endpoint:

deno run --allow-net --allow-env --allow-read main.tsx

Then visit:

http://localhost:8000/search/test?q=How%20to%20use%20Groq%20API

Look for:

{
  "metadata": {
    "strategy": "transformers-local-onnx",
    "localModel": true,
    "modelPath": "/path/to/models/all-MiniLM-L6-v2",
    "timings": {
      "queryEmbedding": 25,
      "total": 65
    }
  }
}

Cause: Model files not found
Fix: Make sure you ran the download script:

cd search/models
./download-model.sh

Verify files exist:

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

You should see model.onnx (~23MB)

Cause: Wrong import path
Fix: Check search/index.ts has the correct import:

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

Cause: Using wrong strategy (auto-download version)
Fix: Confirm the import is transformers-local-onnx.ts, not transformers-cosine.ts

Edit search/transformers-local-onnx.ts:

const USE_QUANTIZED = true;  // Change from false

Benefits:

• Smaller: ~6MB vs ~23MB
• Faster: ~15-20ms vs ~25-30ms per query
• Slightly less accurate: ~57 vs ~58 MTEB score

If you were using a different strategy before, regenerate embeddings:

GET /cache/recalculate?force=true

This ensures all page embeddings use the same model.

Result: ~2-3x faster, no network dependency! 🚀

• Full Setup Guide: search/models/SETUP.md
• Strategy Comparison: search/STRATEGY-COMPARISON.md
• Search Module Docs: search/README.md
• Main Docs: README.md (see "Search" section)

1. Check search/models/SETUP.md for detailed instructions
2. Read search/STRATEGY-COMPARISON.md to compare all strategies
3. Run search/models/test-local-model.ts to verify your setup
4. Check the console for error messages with helpful hints