Semantic Search
summary
Shows that pre-trained transformer embeddings plus a vector database replace messy language-specific hacks by encoding meaning, with a multilingual embedding model claimed to match OpenAI and supporting cross-lingual search across 100+ languages. Describes pipeline: BERT-like encoder + pooling → cosine similarity, reranker, and on-premise deployment for millisecond latency and data control. Trains with contrastive loss using 900M English triplets and 450M non-English triplets, reports cleaner results (Toronto weather demo) and concrete gains versus lexical search.
AI-written summary · about
chapters (8)
- 0:00Intro
- 1:49Strong NLP Performance
- 2:42The Issues with Lexical Search
- 6:58Multilingual Lexical Search
- 8:00Semantic Search
- 13:31How to Train Embedding Models
- 18:29Language Bias
- 23:00Conclusion
related talks
30mSemantic Search and Neural Information Retrieval
39mAI-powered Semantic Search; A story of broken promises?
29mEnabling the world to Search by Meaning
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
41mSemantic Multimodal Multilingual Similarity Engine
32mBuilding semantic search using Deep Learning
45mHybrid search: Greater than the sum of its parts?
13mA Beginner’s Guide to Semantic Search