Building a Scalable Recommender System with Two-Tower Models and ANN Indexes | by Gaurav Kandel | Mar, 2025


Once the item embeddings are computed offline, they must be stored in a highly optimized Approximate Nearest Neighbor (ANN) index to allow for fast retrieval. ANN indexes come in two main flavors:

These are highly optimized libraries designed for fast nearest-neighbor search:

  • FAISS (Facebook AI Similarity Search) — Optimized for large-scale search.
  • ScaNN (Scalable Nearest Neighbors by Google) — Efficient and fast.

👉 Use Case: When speed is the highest priority, such as real-time product recommendations.

These are full-fledged databases that store embeddings along with metadata and provide advanced filtering and indexing:

  • Qdrant — Open-source and production-ready.
  • Hopsworks — Integrates with ML feature stores.
  • MongoDB (with vector search) — Combines vector search with traditional database features.

👉 Use Case: When we need long-term storage, metadata filtering, and scalable querying (e.g., filtering by category or price).

Recent Articles

Related Stories

Leave A Reply

Please enter your comment!
Please enter your name here