How Suning Built a High‑Performance Search Engine with AI‑Driven Segmentation and Ranking

In previous issues we explored Suning's big data initiatives; this article introduces Suning's search platform.

Suning IT headquarters' search team consists of three global teams: Nanjing Search Engine R&D Center, Beijing Search Engine R&D Center, and the Silicon Valley Search Engine R&D Center, which collaborate through audio‑video, network communication, and global development platforms.

1. Search Engine Product Construction

1. Segmentation

The segmentation module implements dictionary‑based tokenization and trigram probabilistic tokenization, optimized for e‑commerce product data. In November 2014, Suning replaced its previous system with a new one that indexes one million effective words and achieves over 98% accuracy, significantly improving precision. Synonym mining produced tens of thousands of product‑related synonym groups, boosting recall.

2. Ranking

Suning's ranking leverages proprietary semantic analysis, using sequence labeling, CRF‑based central word identification, and relational models. By integrating a product knowledge graph and unstructured user behavior data, the system identifies search intent and combines factors such as sales, new arrivals, product popularity, and quality scores to present more accurate results. After several iterations, the classification prediction rate exceeds 97%.

3. No‑Result Handling

Since March 2015, Suning has deployed typo correction and query clustering for no‑result terms, providing keyword recommendations that dramatically improve fault tolerance and product recall. The knowledge‑base module handles over 95% of no‑result traffic, adding roughly 300,000 additional product impressions and greatly enhancing operational efficiency.

Through multiple version upgrades, the average conversion rate rose from 43% to 58% within a year.

2. System Construction

The search system's cluster, completed in 2014, comprises distributed data collection, processing, index creation, retrieval, support, and semantic analysis components. It supports horizontal scaling, handling nearly ten million products with full‑batch processing under two hours and incremental updates within fifteen minutes. System availability exceeds 99.99%, with no major incidents and support for tens of thousands of concurrent users.

Facing explosive SKU growth, the three R&D centers are planning forward‑looking enhancements using big data analytics, natural language processing, voice search, and image recognition to further improve the search engine's usability.