Zhiyi Technology (知衣科技)

Solution

Zhiyi Technology rebuilt its large-scale image-matching workflow on Alibaba Cloud MaxCompute vector search while preserving its own computer- vision models. Social-media images and product images are first cleaned and cropped so only the garment region is passed downstream, which reduces noise from people, backgrounds, and accessories. Zhiyi's self-developed CV model then labels the cropped images and generates feature vectors on GPU clusters. Instead of shipping those vectors into a separate external vector-retrieval stack, the company writes them directly into MaxCompute and uses VECTOR_SEARCH SQL to perform massive offline similarity matching. That lets the business keep the model layer in-house while outsourcing the heavy retrieval layer to managed infrastructure. The new pipeline lowers CPU and memory cost sharply, reduces operational burden, and makes high-frequency full matching practical enough for ongoing merchandising and trend-selection workflows.

Data flow

Product and overseas social-media images are cleaned, cropped to the garment region, and then processed by Zhiyi's CV model on GPU clusters to generate vectors. Those vectors are written into MaxCompute, where VECTOR_SEARCH SQL performs batch similarity matching so merchandising teams can connect products with creator content and trend signals for product-selection decisions.