Vespa.ai’s Personalized Search: Advanced Ranking & Tensor framework
summary
Vespa demonstrates a working e‑commerce demo built in days that combines LLMs (lava, GPT‑4o style prompts) to generate item descriptions and sparse tensor features, indexes dense embeddings with HNSW and BM25 for hybrid retrieval, then applies multi-stage rerank and real-time per-user vectors to update personalization on the fly. Claims include millisecond latency at petabyte scale and beyond 100k QPS, L2‑normalized sparse tensors for user profiles, configurable HNSW params, and chunked ranking to return only top relevant passages.
AI-written summary · about
chapters (10)
- 0:00Introduction and demo
- 1:13Sparse feature engineering
- 9:32Implementing recommendation logic
- 14:29Vespa ranking profiles
- 18:45History and architecture
- 22:43Building blocks of Vespa
- 26:51Advanced ranking framework
- 29:31Multi-stage ranking pipeline
- 31:46Tensor framework formalism
- 33:24Advanced examples and Q&A
related talks
22mTensor arithmetics in search and ranking for Ecommerce.
20mUnderstanding Vespa with a Lucene mindset
23mMachine Learning for Apache Solr, Elasticsearch & Vespa #bbuzz
22mAI-Powered Personalized Discovery with Graph Search and VLM-Enhanced Embeddings
31mAdvanced search for your e-commerce website
26mPersonalizing User Content - The Power of a Vector & Search DB
39mAI-Powered Search in Modern E-Commerce
39mAI-powered Semantic Search; A story of broken promises?