Learning to hybrid search
summary
Uses the Amazon ESCI dataset and Metarank to compare BM25, LambdaMART LTR, sentence-transformer miniLM embeddings and cross-encoders; BM25/LTR sit around ~76 NDCG@5, fine-tuned bi-encoders reach ~78.8, and a fine-tuned cross-encoder scores ~81.3. Builds a hybrid pipeline that ensembles lexical + vector retrievers and a slow but high-quality cross-encoder re-ranker, shows improved relevance but notes production tradeoffs: cross-encoder re-ranking of ~100 candidates costs ~700 ms and GPU/operational expense.
AI-written summary · about
chapters (9)
- 0:00Introduction to hybrid search
- 4:12Baseline BM25 and random ranking
- 9:28BM25 improvements with LambdaMART
- 12:37Neural search with B encoders
- 17:58Fine-tuning embeddings for relevance
- 27:40Cross encoders for re-ranking
- 33:01ESCI dataset and TREC competition
- 34:47Key takeaways and conclusions
- 35:27Q&A session
related talks
33mAI-Powered Search in Modern E-Commerce
01h 30mBuilding a Search Engine
44mRediscover your keyword search: Expand, Enrich and Rewrite
22mAgentic Hybrid Retrieval: Orchestrating LLM Agents, Hybrid Search & Explainability in OpenSearch
49mHybrid search on hybrid models, at scale
38mNatural Language Search-Lexical and Semantic
01h 08mGoing multi-modal
09mFine-tuning embeddings 4 vector search