The Challenge of Industrial-Scale Item Management

Managing product catalogs at the scale of JD.com—which serves over 700 million users and handles tens of billions of SKUs—requires more than traditional database management. The platform faces three primary hurdles: the rapid emergence of new product concepts, the need for high-quality knowledge production across massive datasets, and the requirement to serve diverse downstream applications like search and recommendation.

The Oxygen AIIC Architecture

To solve these, JD developed the Oxygen AI Item Center (Oxygen AIIC), which relies on four core pillars:

• Collaborative Ontology Engineering: The system uses a human-AI loop to manage an evolving ontology containing millions of entries, ensuring the taxonomy remains agile and accurate.
• Semantic Search then Discrimination (S2D): This architecture is the engine of the platform. It first performs a semantic search to retrieve relevant context, followed by a discrimination phase to verify and structure the data. This approach enables high-throughput processing suitable for tens of billions of items.
• Self-Evolving Models: The platform utilizes LLMs and VLMs that improve iteratively. By maintaining a stable, controllable feedback loop, the system achieves 94.2% precision and 82.8% recall in knowledge extraction.
• Unified Item Tunnel: A centralized data and service hub that standardizes how item knowledge is ingested and distributed across the organization.

Operational Impact

The platform currently processes hundreds of millions of item updates daily on Huawei Ascend NPUs. The integration of Oxygen AIIC into core business functions has yielded significant improvements: search-traffic coverage has reached 80.4%, item-information quality issues have decreased by 37%, and the automated fill rate for core product attributes during listing now exceeds 80%.