Revolutionizing Search with Power of Neural Sparse and Dense Vectors
summary
Hybrid search combines 768-dim embeddings (BERT-style dense vectors) with neural-sparse token-weight vectors and runs them in parallel inside OpenSearch ML plugin, storing both as k-NN vectors and an inverted index. The demo indexes an e‑commerce phone dataset with two embeddings per document, merges and re-ranks results, and shows fewer hallucinations and better relevance than BM25 or dense-only search, at the cost of extra indexing/latency, score-normalization and parameter tuning. Implementation works end-to-end but requires tuning for scale and exact-match guarantees.
AI-written summary · about
chapters (11)
- 0:00Hybrid search with neural vectors
- 2:11Traditional search limitations
- 8:52Dense vector embeddings
- 17:59Dense vectors drawbacks
- 20:51Neural sparse vectors role
- 26:51Hybrid search benefits
- 31:06Hybrid search in practice
- 34:35OpenSearch hybrid queries
- 36:01Hybrid search demo
- 40:26Key takeaways and future
- 41:53Q&A on performance
related talks
22mVector Search Made Simple: Getting Started With OpenSearch for AI Applications
44mRediscover your keyword search: Expand, Enrich and Rewrite
38mNatural Language Search-Lexical and Semantic
48mUnlocking Next-gen Semantic Search With OpenSearch 3.x
28mUse OpenSearch as Your Vector Search Database To Optimize GenAI Accuracy!
30mVector Search Made Simple: Getting Started with OpenSearch for AI Applications
44mVector Search Isn’t Enough
27mMachine Learning for Advanced AI Search