A Case on Building a Multilingual Business Matching System With OpenSearchService
summary
Migrates a legacy SQL search that suffers 10s latencies to a FastAPI front end backed by Amazon OpenSearch Service, replacing brittle code-based synonyms with a searchable lexicon index that updates in real time. Uses Kuromoji-like Japanese tokenization, stores translated text in multi fields, and runs a hybrid of BM25 keyword scoring plus Cohere embeddings k-NN on nested child chunks; score_mode=max picks the best business-unit match. The result eliminates profile dilution, preserves exact-term precision, and enables live synonym and translation updates.
AI-written summary · about
chapters (8)
- 0:00Multilingual business matching intro
- 1:49Business matching challenges
- 3:40Legacy SQL limitations
- 5:38Architecture strategy
- 7:33Language barriers with ML
- 9:45Solving noise problems using vector search chunking
- 11:41Hybrid search methods
- 12:31Solutions recap
related talks
41mText to Context: How We Introduced Hybrid Search
30mSearch on Steroids — A Story of Semantics
26mAdvancing Multi-Modal Search Capabilities in Search Pipeline
25mSemantic Search
12mMaximize media search scalability & performance w/
01h 30mBuilding a Search Engine
18mAdvancing Search at Uber with OpenSearch
47mWhy, What, When and How of Running Diff Types of Sear...