The Architecture of Industrial-Scale Knowledge Production

JD.com developed the Oxygen AI Item Center (Oxygen AIIC) to manage item data for over 700 million users and tens of billions of SKUs. The system addresses the challenge of high-volume, dynamic e-commerce data through four primary pillars:

• Human-AI Collaborative Ontology: An agile system that supports the dynamic evolution of an ontology containing millions of entries, ensuring the platform adapts to fast-emerging market concepts.
• S2D (Semantic Search then Discrimination) Architecture: This identification architecture optimizes throughput for massive SKU volumes. By decoupling the search phase from the discrimination phase, the system achieves the scalability required to process hundreds of millions of daily updates.
• Self-Evolving LLMs/VLMs: The core models are designed for stable, controllable improvement, resulting in a knowledge production pipeline that maintains 94.2% precision and 82.8% recall.
• Unified Item Tunnel: A centralized data and service hub that streamlines integration across the organization.

Operational Impact and Performance

Oxygen AIIC is deployed on Huawei Ascend NPUs, enabling high-throughput processing of hundreds of millions of item updates daily. The platform has successfully integrated into core business workflows, including search, recommendation, and category planning. Key performance metrics include:

• Search-traffic coverage: Reached 80.4%.
• Data Quality: Item-information quality issues decreased by 37%.
• Operational Efficiency: The automated fill rate for core attributes during item listing now exceeds 80%.

By moving from manual or rule-based management to an LLM/VLM-centric approach, JD.com has demonstrated that industrial-scale item understanding can be both automated and highly reliable, significantly reducing management costs while improving consumer experience.