Lexical, Vector & Hybrid Search with Elasticsearch
summary
Explains how Elasticsearch (Lucene) uses Okapi BM25 for lexical ranking and when that fails due to vocabulary mismatch; demonstrates kNN vector search with HNSW graphs, cosine/L2/dot‑product choices and embeddings (typical dims 256–1024), warns of higher memory and slower indexing, and shows compression options like BBQ quantization. Shows two hybrid blends: weighted linear combination (boost tuning) and Reciprocal Rank Fusion (k≈60), plus reranker and RAG patterns to ground LLMs with retrieved context.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:531st search engine
- 3:37Lexical search
- 12:57Vector search
- 23:14Hybrid search
- 27:22Reranking models
- 28:59AI search
- 35:40Summary
- 36:16Outro
related talks
42mResearch in the Age of AI
01h 09mVector Search: Ask Me Anything!
41mText to Context: How We Introduced Hybrid Search
30mSearch on Steroids — A Story of Semantics
38mSearch: A New Era
46mNatural Language & Search, Improve your Search w/ Vector capabilities
13mA Beginner’s Guide to Semantic Search
44mVector Search Isn’t Enough