Search on Steroids — A Story of Semantics
summary
A company built a hybrid search by using Elasticsearch as a vector database combined with traditional lexical search to index about 3,500 products from 70 brands. Items and queries are vectorized with open models from Hugging Face, results are merged with Reciprocal Rank Fusion outside Elasticsearch, and the hybrid search preserves exact-keyword hits while returning semantically related items; this reduced rapid repeated queries and increased conversions. They accept trade-offs of extra compute, indexing/vectorization cost, higher latency and multilingual/data-quality work, and they also use a RAG layer with OpenAI to power chatbots using the filtered vector context.
AI-written summary · about
chapters (8)
- 0:00Advanced search introduction
- 2:08Lexical search limits
- 5:19Vector search principle
- 12:47Hybrid search benefits
- 16:48Solution limits and costs
- 20:17Advanced applications and RAG
- 27:14User impact summary
- 27:56Q&A
related talks
41mText to Context: How We Introduced Hybrid Search
31mAdvanced search for your e-commerce website
37mLexical, Vector & Hybrid Search with Elasticsearch
32mBuilding semantic search using Deep Learning
01h 09mVector Search: Ask Me Anything!
44mVector Search Isn’t Enough
47mWhy, What, When and How of Running Diff Types of Sear...
01h 48mInformation Retrieval from the Ground Up