From text search & recommendation to ads & online dating
summary
Modern ANN libraries lose many relevant results when metadata filters are applied after retrieval; Vespa modifies HNSW to support eligibility lists so ANN search only considers items that pass filters, supports incremental adds/removes for immediate indexing, and measures competitive recall/throughput on ann-benchmarks while serving hundreds of thousands qps in production for search, ads, dating and shopping. If a filter is very strong Vespa falls back to exact scan over eligible items to keep cost practical.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 0:55What is Vespa
- 2:22Representations
- 3:48Finding nearest neighbors
- 4:37Approximate solutions
- 5:38NNN Benchmarks
- 8:49Metadata
- 9:16Text search
- 9:47Queries
- 10:32Filtering
- 11:48Recommendation
- 13:30Ads
- 14:38Online dating
- 15:34Shopping
- 16:22Local Business
- 17:18Postfiltering
- 18:28Eligibility lists
- 19:20Configurable threshold
- 19:40Summary
- 20:12Use cases
- 21:15Conclusion
- 21:44Questions
related talks
40mVespa.ai’s Personalized Search: Advanced Ranking & Tensor framework
30mSemantic Search and Neural Information Retrieval
40mUnder the hood of vector search with JVector
20mUnderstanding Vespa with a Lucene mindset
42mResearch in the Age of AI
32mBudget Friendly Semantic Search With Neural Sparse Search
26mPersonalizing User Content - The Power of a Vector & Search DB
36mWhen Semantic Search Breaks in your RAG systems: RAM Walls, Silent Failures, and the Architecture Decisions