How MaxCompute Vector Search Bridges Products and Overseas Social Media for Real‑Time Cross‑Border Selection

Cross‑border e‑commerce merchants need early signals of product popularity beyond in‑store sales; overseas social‑media exposure often appears earlier, but manual keyword browsing cannot reliably capture it due to limited coverage and high labor cost.

To turn this signal into a selection decision, the problem is framed as an image‑retrieval task: convert product images and influencer photos from overseas platforms into feature vectors and match them at scale. Three engineering challenges are identified: (1) locating the clothing subject in mixed‑content images, (2) handling the compute pressure of full‑batch cross‑matching (≈10 M product vectors vs. ≈2 M social‑media vectors, each 512‑dimensional), and (3) keeping retrieval cost low enough to support daily large‑scale re‑runs.

In partnership with Alibaba Cloud’s big‑data AI platform, the workflow was rebuilt. After cleaning raw social‑media posts, both product and influencer images undergo clothing‑area detection and cropping. The cropped images are processed by ZhiYi’s self‑developed CV model on a GPU cluster to generate labeled feature vectors, preserving the company’s model assets. Batch similarity matching is then performed entirely on MaxCompute using the built‑in VECTOR_SEARCH SQL function, which automatically builds and maintains indexes. This replaces the previous Proxima engine that required separate index management and higher operational overhead.

With the new pipeline, merchants can directly view which influencers have featured a given product and assess competitor exposure, turning off‑platform social signals into a quantifiable metric for product selection. The presence of influencer‑generated content now complements internal sales data, providing a stronger, earlier indication of market demand.

Performance testing on identical data volumes (≈10 M product vectors and ≈2 M social‑media vectors) shows that MaxCompute’s offline vector search reduces CPU consumption by about 73% and memory consumption by about 93% compared with the historical Proxima solution. The lower resource cost allows larger content pools or higher re‑run frequencies within the same budget, eliminating the trade‑off between pool size and data freshness.

According to the ZhiYi technical team, the collaboration eliminated the previous bottleneck of retrieval cost, letting them focus on improving clothing feature extraction while the cloud platform handles large‑scale indexing and batch retrieval. Future plans include extending the capability to more e‑commerce platforms and scaling to even larger social‑media image collections.