From Embeddings To Index: A Practitioner's Guide To Domain-Adapted Neural Search
summary
Off-the-shelf embeddings lose most relevance on private jargon, costing 40–60% search relevance versus 1–5% ANN indexing loss; the practical pipeline is vocabulary extension → domain warm‑up → dictionary pre‑training → multi‑objective fine‑tuning, with each stage measured for impact. Tokenizer edits plus synthetic term‑definition corpora revive dead logits; sparse (SPL ADE style) vs dense projections trade relevance for storage cost; index maintenance uses reconstruction‑error tracking with a candidate quantizer to decide between lazy adapt or full rebuild.
AI-written summary · about
chapters (8)
- 0:00Neural search adaptation intro
- 1:08Neural search problem
- 6:08Fine-tuning limits for domain adaptation
- 12:49Embedding model adaptation steps
- 27:57Production-ready embeddings
- 32:58Embedding adaptation impact on indices
- 34:48Scaling index efficiency
- 40:29Key takeaways & Q&A
related talks
45mBehind the hype: managing billion-scale embeddings in Elasticsearch and OpenSearch
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
32mBudget Friendly Semantic Search With Neural Sparse Search
30mSemantic Search and Neural Information Retrieval
13mThe State of AI Powered Search and Retrieval
13mA Beginner’s Guide to Semantic Search
46mNatural Language & Search, Improve your Search w/ Vector capabilities
38mSearch: A New Era