Chan Long, Chao Liu, Chaofan Chen, Chaohui Dong, Chunyuan Guo, Danping Liu, Debin Liu, Deping Xiang, Fulai Xu, Guangyue Liu, Hao Li, Huichun Hu, Jian Yang, Jianan Wang, Jianbo Zhao, Jiaoyang Li, Jiaxing Wang, Jinglong Li, Jinjin Guo, Jun Fang, Jun Liu, Kai Zhou, Li Wang, Lili Gao, Liying Chen, Luning Yang, Mengdi Zhou, Pengzhang Liu, Qi Lv, Qianyun Wang, Qixia Jiang, Ruyue Li, Shimu Liang, Shuxing Wang, Sijie Zhang, Siqi Li, Tianhao Gao, Wang Ke, Weihu Huang, Wencan Lai, Wenjie Zhang, Xiaohui Zhang, Xiaojing Dong, Ya Liu, Yifeng Zhang, Yixiang Wang, Yongtai Zhang, Yongyi Liao, Zhaoru Chen, Zhen Chen, Zhiyong Ma, Zhiyuan Liu, Zhongwei Liu, Ziyan Xing
5 min
As e-commerce platforms grow to manage tens of billions of SKUs, traditional manual or rule-based item management systems struggle to keep pace with fast-emerging market trends, semantic gaps in user queries, and the high cost of maintaining structured product data. This paper addresses how to build an industrial-scale, LLM/VLM-centric infrastructure that can automate high-quality item knowledge production, management, and service across diverse business scenarios.
The authors introduce the Oxygen AI Item Center (Oxygen AIIC), which is built on four pillars:
Oxygen AIIC has successfully scaled to process hundreds of millions of item updates daily, accumulating hundreds of billions of knowledge assets. The platform demonstrates a 94.2% precision and 82.8% recall in knowledge production. In production, this has led to tangible business improvements: search-traffic coverage reached 80.4%, item-information quality issues decreased by 37%, and the automated fill rate for core attributes exceeded 80%. Furthermore, the system reduced decision cycles for category planning from weeks to days.
This work provides a blueprint for deploying large-scale generative AI in mission-critical industrial environments. By demonstrating how to combine expert-in-the-loop oversight with automated model-driven pipelines, the authors show that LLMs can move beyond general-purpose tasks to become reliable, high-throughput infrastructure for complex, real-world data management.
JD$.$com, one of the world's largest e-commerce platforms, serves over 700 million active users and millions of merchants, with a catalog of tens of billions of SKUs. At this scale, high-quality, structured item knowledge underpins a better consumer experience, lower management costs, and higher operational efficiency-yet producing and serving it poses three industrial-scale challenges: fast-emerging concepts, high-quality knowledge production for massive SKUs, and diverse downstream requirements. To address these challenges, we present the JD Oxygen AI Item Center (Oxygen AIIC), an industrial-scale platform built on LLMs/VLMs for item-knowledge production and service. Oxygen AIIC is built around four core pillars: (i) ontology engineering driven by efficient human-AI collaboration, which supports the dynamic evolution and agile expansion of an ontology with millions of entries; (ii) a "Semantic Search then Discrimination"(S2D) knowledge identification architecture that, combined with throughput improvement strategies, enables scalable, extensible, and high-throughput AI Item Library production for tens of billions of SKUs; (iii) self-evolving item-understanding LLMs/VLMs that improve in a stable and controllable manner, enabling knowledge production with 94.2% precision and 82.8% recall; and (iv) a unified item tunnel that serves as the data and service hub. Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day on Huawei Ascend NPUs. It has accumulated hundreds of billions of item-knowledge assets. Deployed across core business scenarios-including search, recommendation, operations, category planning-Oxygen AIIC has delivered measurable gains at scale. Search-traffic coverage reaches 80.4%, item-information quality issues drop by 37%, the automated fill rate of core attributes during item listing exceeds 80%.
Sam: That's a fair concern, and the paper addresses it directly. They build in what they call a "self-evolution" loop. The system continuously tests its own predictions against real-world signals — things like whether users actually click on and buy the items it categorizes. When a new trend starts emerging and the system's confidence drops, it flags those cases for human review. Experts then update the ontology, and that updated map feeds back into the AI. So the humans aren't doing the repetitive cataloging work anymore — they're acting more like editors, steering the system toward accuracy when it encounters something genuinely new.
Alex: So the human role shifts from doing the work to supervising the work.
Sam: Exactly. And that shift is what makes the scale possible. The paper reports over 94% precision in attribute tagging, and that accuracy translated to around 80% of product attributes being filled in automatically — work that would otherwise require an enormous manual effort. Click-through rates also increased, which suggests the improved catalog quality was meaningful to real shoppers.
Alex: One thing I'm curious about — running checks like this across ten billion items sounds computationally expensive. How do they keep up?
Sam: They rely on specialized computer chips called Neural Processing Units, or NPUs. Standard chips are general-purpose — they handle everything from running a spreadsheet to playing a video. NPUs are purpose-built to accelerate the specific kind of math that AI models run. It's the difference between using a kitchen knife for everything versus having a dedicated bread knife. Using NPUs means the system can process the volume of a global marketplace without the checks becoming a bottleneck.
Alex: So the broader lesson here isn't really about building a smarter AI brain. It's about building a better system around the AI — a reliable map, a two-step verification process, a feedback loop with human experts, and the right hardware to run it all at scale.
Sam: That's a precise way to put it. The shift from asking an AI to "know" everything to asking it to "search and verify" against a well-maintained reference map turns out to be a more practical and more reliable approach — at least at this scale. It's a useful reminder that system design often matters as much as model capability.
Alex: Thanks for walking us through that, Sam. And thanks to everyone listening to ResearchPod.