Empowering Customer Decisions with Elasticsearch: From Search to Answer Generation

Elastic Inc. Booth, AWS re:Invent 2024
https://reinvent.awsevents.com/

1. Empowering Customer Decisions with Elasticsearch: From Search to Answer Generation Taisuke Hinata Search Engineer @ Nikkei Inc. December 3, 2024 Elastic Inc. Booth, AWS re:Invent 2024
2. Overview • Nikkei and Our Database Business • Elasticsearch in Our Search • From Search to Answer Generation
3. Nikkei Inc. • Publishing Japan’s Largest Business Newspaper ◦ 3+ million subscribers • Founded in 1876
4. News Database Media Business Index Development Education Broadcasting Publishing Culture Project
5. News Database Media Business Index Development Education Broadcasting Publishing Culture Project
6. Data to Action. Helping our customers make better decisions. https://nkbb.nikkei.co.jp/about/
7. Our Database Business 200+ Media Companies ・ ・ ・ +
8. Our Database Business All contents Content Platform (CPF) 200+ Media Companies ・ ・ ・ + Since 2018
9. Our Database Business All contents Content Platform (CPF) 200+ Media Companies ・ ・ ・ ・ ・ ・ 12 Information Search Services + Since 2018
10. • Business Search Service since 1984 ◦ 750+ media sources ◦ 86M+ companies ◦ 300K+ individuals • Trusted by 70% of Japan’s largest publicly listed companies • Supports decision-making
11. Search is the Key to Our Business Decision Making Support Reliable Contents Search Technology
12. Elasticsearch in Our Search
13. Transition of Search Engines 1999 2007 2014 2023 May 2024 2018 6 8 7 on-premises PanaSearch
14. Our System Architecture Content Platform (CPF) 200+ Media Companies ・ ・ ・ ・ ・ ・ 12 Information Search Services + ETL APIs DB
15. Scale of Our Elasticsearch • 48-node cluster in 3AZs • 190M docs • +15K / day • Performance ◦ 200 QPS, ≤200ms Coordinating Only Nodes (3) Data Nodes (42) Master Nodes (3) ・ ・ ・
16. One of Japan’s Largest Elasticsearch https://www.elastic.co/jp/customers/nikkei-1 https://www.acroquest.co.jp/archives/15788/
17. Powers Our All Search Features Query Suggest Search Results Display Body
18. Query Suggest Search Results Display Body Keywords are suggested Powers Our All Search Features
19. Query Suggest Search Results Display Body Results are aggregated by theme Powers Our All Search Features
20. Query Suggest Search Results Display Body Matching text is highlighted Powers Our All Search Features
21. Query Suggest Search Results Display Body Related articles are displayed together Powers Our All Search Features
22. From Search to Answer Generation
24. Steps to Knowledge: Traditional Search Question Query Select Read Answer What is TOYOTA’s business situation in Poland? 🤔 TOYOTA AND Poland AND expansion TOYOTA TOYOTA TOYOTA
25. Question Query Select Read Answer What is TOYOTA’s business situation in Poland? 🤔 TOYOTA TOYOTA TOYOTA TOYOTA AND Poland AND expansion Steps to Knowledge: Generative AI
26. How to Resolve Question DB Retriever Answer Ask doc Summarize Query Question & Docs Generative AI • Retrieval Augmented Generation (RAG) ・・・
27. Elasticsearch Has the Edge in RAG Elasticsearch Most Vector Database Some Vector Database Create Vector Embeddings Store & Search Vector Embeddings Search Analytics Hybrid Search (text + vector) Ingest Tools(Web Crawler, connectors, Beats, Agent, API framework) Playground for testing RAG Autocomplete Choice & Flexibility of embedding models
28. Our New RAG Service with Elasticsearch
29. Company Analysis Market Analysis Economic Trends Analysis Trend Investigation Proposal Support Report Creation Gain insights into a company’s external environment and initiatives. Analyze industries with PEST analysis and market trends. Explain Nikkei Index trends and analyze CPI changes. Predict business trends and consumer behavior shifts. Create proposal stories and business roadmaps. Generate reports based on specific questions. Please enter your question.
33. Current Architecture Question 5 Sub Queries Results Article Answer Gemini Flash text-embedding -3-large Top 30 results Fetches 1000 docs (200 *5) Gemini Pro knn mseach text-embedding -3-large Chunks
34. Next Steps • Hybrid Search (Text + Vector) ◦ Now, combining Japanese text search with vector search reduces accuracy. • Japanese Sparse Vector Model ◦ Dense vectorization (1536 dim) of 190M documents requires 3.0 TB (2.35x). ◦ Elastic provides ELSER as a sparse vector model, optimized for English.
35. Summary • Nikkei, originally a newspaper company, provides cross-search services across content from 200+ partner companies. • Search is at the core of our business, and it is powered by Elasticsearch. • In the era of generative AI, we continue to leverage Elasticsearch to accelerate customer decision-making!