A Practical Approach To Semantic Search
summary
Compares lexical vs semantic on Amazon-ESCI product data and shows that when both are fine-tuned they reach similar NDCG but semantic search has much lower precision, which is unacceptable for product search; only 28.6% overlap of returned items on average, with lexical winning on short exact queries and semantic winning on long, complex natural-language queries. Implements encoders via ONNX (indexing becomes 7x slower with ONNX) then deploys a Triton GPU service across dozens of nodes; L-fusion raises NDCG but cuts precision, and LM-based labeling finds false negatives and matches human labels.
AI-written summary · about
chapters (10)
- 0:00Introduction to search
- 2:04Search architecture
- 5:35Choosing evaluation metrics
- 7:58Performance comparisons
- 10:41Query-type performance
- 15:32Integrating search strategies
- 21:48Data and annotation
- 24:54Operational challenges
- 27:48ROI and conclusion
- 30:24Q&A session
related talks
13mA Beginner’s Guide to Semantic Search
32mBuilding semantic search using Deep Learning
22mThe Capabilities of Semantic Search
32mBudget Friendly Semantic Search With Neural Sparse Search
30mSemantic Search and Neural Information Retrieval
37mLexical, Vector & Hybrid Search with Elasticsearch
30mSearch on Steroids — A Story of Semantics
37mSimilarity Search for Product Matching @ Semantics3