How We Built AI-Powered Search That Runs Entirely in Your Browser

Embedding Pipeline

// Load model once, reuse forever
const pipeline = await pipeline('feature-extraction', 
  'Xenova/all-MiniLM-L6-v2');

// Generate embeddings locally
const embeddings = await pipeline(text, {
  pooling: 'mean',
  normalize: true
});

The entire pipeline runs in WebAssembly. No network requests after initial model download.

Vector Search Implementation

// Cosine similarity in pure JavaScript
function cosineSimilarity(a, b) {
  let dotProduct = 0;
  let normA = 0;
  let normB = 0;
  
  for (let i = 0; i < a.length; i++) {
    dotProduct += a[i] * b[i];
    normA += a[i] * a[i];
    normB += b[i] * b[i];
  }
  
  return dotProduct / (Math.sqrt(normA) * Math.sqrt(normB));
}

// Search across 10,000 embeddings in ~50ms
function search(queryEmbedding, storedEmbeddings) {
  return storedEmbeddings
    .map(item => ({
      ...item,
      score: cosineSimilarity(queryEmbedding, item.embedding)
    }))
    .filter(item => item.score > 0.7)
    .sort((a, b) => b.score - a.score);
}