Personalizing User Content - The Power of a Vector & Search DB
summary
Taboola replaces offline candidate lists with a Vespa-based vector search that ranks millions of items into a few thousand candidates in ~300ms, using Debezium + Kafka for near‑real‑time writes and an async pipeline that rebuilds item vectors every few minutes. Traffic peaks reach hundreds of thousands of requests/sec across 10,000+ publishers; filters originally consume ~60% of cost, so documents are stored as campaign matrices to cut CPU/memory and achieve 60% resource savings. Deployment uses multi-cluster canary rollout and vector-version metadata to avoid global bad-rollouts.
AI-written summary · about
chapters (16)
- 0:00Introduction to recommendation systems
- 1:31What is Kadish in recommendations
- 3:40Offline Kadish limitations
- 6:12Moving Kadish online with vector search
- 9:02How deep learning models work
- 10:30Generating item vectors with two-tower models
- 11:57Vector DB integration in recommendation flow
- 12:40Beyond clicks: modeling conversions
- 13:44Choosing Vespa for vector search
- 15:08Real-time vector updates and data pipelines
- 17:40Challenges with vector DB updates
- 18:44Multi-cluster rollout strategy
- 20:12Managing vector versioning
- 21:59Cost optimization in vector search
- 23:26Matrix-based filtering in Vespa
- 24:30Summary of vector search benefits
related talks
31mAdvanced search for your e-commerce website
40mVespa.ai’s Personalized Search: Advanced Ranking & Tensor framework
41mPractical Vector Search Configuration, Tuning, RAG, and the Road Ahead
31mAgents Need More than Vector Search
50mVector Ops: run vector embedding-powered apps in prod
18m: Workforce Efficiency with AI-Driven Vector Search
01h 09mVector Search: Ask Me Anything!
42mDocument Vectors in the Wild