Hybrid search on hybrid models, at scale
summary
Extensible hybrid search runs on Elasticsearch multi-index architecture with OpenVINO inference, Spring Boot microservices and Kubernetes; Whisper transcriptions feed a sentence-transformer model (124-dim, 512-token limit) and chunking moves from speaker turns to optimized 512‑token and semantic chunks. Indexing shrinks from ~850M short-turn vectors to ~55 million with int8 quantization (~55 GB RAM). Hybrid ranking uses script_score normalization and aggregation across indices as an alternative to reciprocal rank fusion, retrieving up to 1,000 candidates for downstream reranking.
AI-written summary · about
chapters (6)
- 0:00Introduction and context
- 12:37Lexical search limitations
- 20:09Semantic search challenges
- 32:22Architecture and flexibility
- 42:24Key takeaways and evaluation
- 43:29Q&A session
related talks
37mEnhancing AI and LLM Systems with Hybrid Search
41mGenAI, vector/semantic/hybrid search, RRF, NLP, (...) and other buzzwords
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
44mRediscover your keyword search: Expand, Enrich and Rewrite
22mAgentic Hybrid Retrieval: Orchestrating LLM Agents, Hybrid Search & Explainability in OpenSearch
47mBoosting AI With OpenSearch: Vector Stores, Similarity Search, and Retrieval-Augme...
21mBuilding RAG With OpenSearch ML Plugins: From PDFs To Voice-Enabled Search
44mVector Search Isn’t Enough