Unlocking Smarter Typeahead Search
summary
Shows a reproducible hybrid: a low‑latency prefix table plus semantic retrieval built with open source tools and deployed on GPU, encoding 200 million titles with all‑MiniLM‑L12‑v2 and searching with FAISS IBF‑PQ compressed vectors. Claims include coverage and relevance gains, a cost cut by compressing embeddings to fit on a T4, and measured business impact — suggestion engagement lifts up to 37% while serving millions of requests under 62 ms.
AI-written summary · about
chapters (22)
- 0:00Smarter typeahead search intro
- 1:29Typo and discovery challenges
- 2:33Prefix search limitations
- 4:39Prefix suggestions fail
- 7:07Semantic search for coverage
- 8:10Semantic search encoding
- 10:15Encoder choice for speed
- 12:25Vector index options
- 14:09Inverted file index clustering
- 15:36HNSW graph for nearest neighbor
- 17:25Vector compression with PQ
- 19:55Cost savings from compression
- 20:16Hybrid prefix and semantic
- 21:44Semantic search architecture
- 23:09Suggestion coverage results
- 25:01Encoder impact on quality
- 26:30Full system performance
- 27:13Hybrid search summary
- 27:58Q&A on sources and history
- 30:06Q&A on prefix updates
- 31:52Q&A on user satisfaction
- 34:41Q&A on cancellation and efficiency
related talks
32mRetrieval Augmentation and Semantic Search at Scale
31mAdvanced search for your e-commerce website
01h 30mBuilding a Search Engine
48mUnlocking Next-gen Semantic Search With OpenSearch 3.x
41mText to Context: How We Introduced Hybrid Search
46mNatural Language & Search, Improve your Search w/ Vector capabilities
40mVespa.ai’s Personalized Search: Advanced Ranking & Tensor framework
24mVector Search at Uber