Deep Dive: Tools Behind Uzum Market's Search
Uzum Market
Uzum Market's search team shares insights on their retrieval and ranking infrastructure, including a custom speller, personalized suggestions, and synonym management. They highlight how they improved typo correction with contextual features and A/B tested changes to boost conversion.
The team at Uzum Market, led by Dmitry Shipilov, detailed the components powering their search engine. They migrated inference from the ElasticSearch Learning to Rank (LTR) plugin to a dedicated ranking service that accepts Open Neural Network Exchange (ONNX) models, enabling more flexibility. A speller service corrects typos in queries, but initial text-only features led to inaccuracies, for example, correcting 'стуль' to 'сталь' instead of 'стул'. They trained a gradient boosting ranker on annotated data with additional features like query frequency and conversion, reducing incorrect corrections by 50 percent and increasing correct fixes by 8 percent. They also implemented personalized suggestions using gradient boosting with features like gender and spending patterns, improving engagement metrics. For synonyms, they faced issues with an uncontrolled synonym base causing query explosion, and they are working on cleaning up unused and harmful synonyms.
- Abbreviations
- LTR = Learning to Rank — обучение ранжированию
- ONNX = Open Neural Network Exchange — открытый формат обмена нейронными сетями
- LLM = Large Language Model — большая языковая модель
- ML = Machine Learning — машинное обучение
- NDCG = Normalized Discounted Cumulative Gain — нормированный дисконтированный совокупный выигрыш
- PR = Pull Request — запрос на слияние изменений
Source: Habr — хаб ML —
original