JD.com — often described as the "Amazon of China" — has published technical details about the Oxygen AI Item Center (Oxygen AIIC), the system it uses to manage its very large inventory. JD serves 700 million users and works with millions of merchants; its catalog contains tens of billions of SKUs. Oxygen AIIC is a core component for tracking and managing that inventory.
Capabilities at a glance
According to JD's paper, Oxygen AIIC covers tens of thousands of JD categories and handles hundreds of millions of item updates per day on Huawei Ascend NPUs. The design combines deep learning models with structured engineering to support industrial‑scale, continuously operating inventory management.
The four key elements of Oxygen AIIC
- Ontology engineering with human–AI collaboration
- Experts distill industry knowledge while algorithms learn from that material to scale ontology construction and drive continuous evolution.
- "Semantic Search then Discrimination"
- In the first stage, the dynamically evolving ontology is externalized as a separate ontology knowledge base, enabling continuous updates without retraining models. In the discrimination stage, the model only decides whether an item matches the retrieved ontology entries. This arrangement reduces task complexity, mitigates model hallucination, and improves generalization as the ontology evolves.
- Self‑evolving item‑understanding LLMs/VLMs
- The system uses incremental learning and model self‑evolution to fill targeted knowledge gaps and reduce catastrophic forgetting. The core approach builds on a robust multi‑task foundation, develops lightweight "expert modules" for incremental requirements, and dynamically integrates them into the expert pool to enable agile capability expansion.
- "Unified item tunnel"
- This is the main interface between Oxygen AIIC and JD's business applications. It supports daily, minute‑level, and second‑level production and distribution pipelines while preserving data consistency.
Technical and strategic considerations
- JD notes two primary technical challenges during large‑scale deployment: model training and inference on Huawei Ascend NPUs, and efficient utilization of compute resources. The platform's reliance on Chinese compute hardware also aligns with broader efforts toward technology sovereignty in China.
Why this matters
- Oxygen AIIC illustrates how modern AI tools can be embedded into back‑office functions such as inventory management, enabling companies to operate at much larger scales and to self‑update and learn with reduced human oversight. Systems like this support continuous operation and rapid adaptation to changes in catalog content and business conditions.
Source
The technical details are documented in "JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial‑Scale LLM/VLM‑Centric Solution for Item Understanding, Management, and Applications" (arXiv).
