The Invasion of Transformers - Boosting Search with Latest NLP
summary
Claims that transformer-backed search moves from lab to production with measurable gains using dense passage retrieval (DPR) plus Approx. Nearest Neighbour (ANN) / HNSW to keep latency constant. On a 100k‑document corpus naive reader QA takes ~30+ hours, adding a retriever cuts end‑to‑end to below one second; benchmark shows DPR+HNSW outperforms BM25 on accuracy while yielding query throughput between DPR exact (≈1.5 QPS) and Elasticsearch (≈80 QPS). The open source Haystack framework implements modular retriever/reader pipelines, vector stores and ANN tuning for production use.
AI-written summary · about
chapters (12)
- 0:00Introduction to transformers in search
- 2:54State of search today
- 7:29Transformers invading search
- 9:16Document retrieval methods
- 12:24Question answering with transformers
- 14:52Demo of QA in action
- 16:22Scaling challenges in production
- 19:57Dense retrieval with DPR
- 24:32Haystack framework overview
- 28:06Building pipelines in Haystack
- 28:49Outlook and future trends
- 29:32Q&A session
related talks
01h 30mBuilding a Search Engine
13mBuilding High-Performance AI Literature Processing
37mEnhancing AI and LLM Systems with Hybrid Search
44mRediscover your keyword search: Expand, Enrich and Rewrite
43mSupercharging your transformers with synthetic query generation and lexical search
37mBuilding Generative AI Applications
45mRevolutionizing Search with Power of Neural Sparse and Dense Vectors
32mSLM-powered retrieval to scale freelancers matching at Malt