Using Approximate Nearest Neighbor Search to Find Similar Products

Jo Kristian Bergum (Vespa) demonstrates ANN-based “similar products” feature on Amazon Fashion dataset (~22K products) using pyvespa.

Key Concepts Demonstrated

ANN with Prefiltering

Combine ANN search with metadata filters in a single query:

nearestNeighbor(image_vector, query_vector) AND !(asin = "X") AND price > 100

Vespa evaluates filters at query time — not post-processing — so filtered recall is maintained.

HNSW Configuration

• max_links_per_node: controls graph connectivity vs. memory tradeoff
• neighbors_to_explore_at_insert: controls index-time recall quality
• Toggle approximate: true/false at query time to switch ANN ↔ exact for A/B testing

Real-Time Partial Updates

Attribute fields (inventory status, price, popularity) can be updated at ~40-50K updates/sec without re-indexing the document. Critical for keeping search index current in retail.

Ranking Integration

ANN results can feed into a multi-signal ranking profile:

12.0 + 23.24 * closeness(field, image_vector) + 12.4 * (1/attribute(popularity))

ANN as first-stage retrieval + business signal re-ranking in one system.

Key Takeaway

The value of Vespa’s ANN isn’t just speed — it’s ANN + filters + real-time updates + ML ranking as a unified system, avoiding the need for separate ANN index + database + re-ranker.