Tensor arithmetics in search and ranking for Ecommerce.
summary
Small 256-d fashion vision model plus Vespa’s ranking engine runs live tensor arithmetic at query time to steer results by color, shape, style and text facets, and demo shows it working over 1.2 million items with interactive sliders that add/subtract control vectors and preserve shape. The model infers ~1,000 embeddings per consumer GPU, Vespa executes SIMD-optimized dot, norm and reduce ops, uses L2 normalization, and supports partial tensor updates (single-field/cell) so embeddings need not be reindexed.
AI-written summary · about
chapters (8)
- 0:00Introduction and background
- 1:04Similar product search demo
- 3:55Modifying tensor embeddings
- 7:05Dynamic ranking with tensors
- 10:37Tensor operations in Vespa
- 13:26Control vectors for personalization
- 16:38Vespa ranking profile mechanics
- 19:52Key takeaways and future
related talks
40mVespa.ai’s Personalized Search: Advanced Ranking & Tensor framework
31mAdvanced search for your e-commerce website
39mAI-Powered Search in Modern E-Commerce
22mMachine Learning: The Engine Behind Modern E-commerce Search
27mData with Purpose: The Power of Embeddings
23mMachine Learning for Apache Solr, Elasticsearch & Vespa #bbuzz
27mMachine Learning for Advanced AI Search
22mAI-Powered Personalized Discovery with Graph Search and VLM-Enhanced Embeddings