Design Visual Search (Search by Image)

Problem

Design Myntra's "search by image" feature where users upload a photo of a clothing item and get visually similar products from the catalogue.

Requirements

Functional:

• User uploads image → system returns top-20 visually similar products
• Support similarity by: overall style, colour, pattern, product type
• Results ranked by visual similarity + availability + price

Non-functional:

• 100M product images in catalogue
• Query latency < 2s
• Index updated within 24h of new product addition

Discussion points

1. Feature extraction: CNN embeddings (ResNet/EfficientNet) → 256-dim vector
2. Approximate Nearest Neighbour search: FAISS, ScaNN, or Pinecone
3. Offline indexing pipeline vs real-time index updates
4. Hybrid ranking: vector similarity + metadata filters (in-stock, price range)
5. Cold start: generating embeddings for 100M images in manageable time