AI/ML pipeline | Cüte Docs

User action → Event tracking → Feature store
                                    ↓
                            ML Pipeline
                                    ↓
                            Predictions → Cache → API

Embedding pipeline

packages/features/items/embeddings.ts

async function generateItemEmbeddings(item: Item) {
  const textEmbedding = await openai.embeddings.create({
    model: "text-embedding-3-small",
    input: `${item.title} ${item.description} ${item.brand}`,
  });

  const imageEmbedding = await generateImageEmbedding(item.mainImageUrl);

  await db.update(items)
    .set({
      textEmbedding: textEmbedding.data[0].embedding,
      imageEmbedding
    })
    .where(eq(items.id, item.id));
}

Similarity search

packages/features/items/queries/similar.ts

async function findSimilarItems(itemId: string, limit = 10) {
  return db.execute(sql`
    SELECT *,
      1 - (text_embedding <=> ${targetEmbedding}) as similarity
    FROM items
    WHERE id != ${itemId}
    ORDER BY text_embedding <=> ${targetEmbedding}
    LIMIT ${limit}
  `);
}

Implementation

Embeddings foundation

Set up pgvector in Supabase, generate embeddings for existing items, and build similarity search API.

Recommendations

Implement user preference tracking, recommendation API endpoints, and "Similar items" on item detail page.

Image recognition

Integrate image classification model, auto-suggest categories on upload, and brand detection from images.

Fraud detection

Add duplicate detection on listing, pricing anomaly alerts, and user behavior scoring.

Checklist

Infrastructure

Enable pgvector extension in Supabase
Add embedding columns to items table
Set up background job for embedding generation
Create vector similarity indexes

Features

Similar items API
User recommendations API
Image auto-tagging
Price suggestion engine
Fraud detection alerts