JD Oxygen AI Item Center (Oxygen AIIC) V1: An Industrial-Scale LLM/VLM-Centric Solution for Item Understanding, Management, and Applications¶

🕒 Published (v1): 2026-06-29 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

JD.com's Oxygen AI Item Center (Oxygen AIIC) is an industrial-scale platform that deploys LLMs and VLMs to produce, manage, and serve structured item knowledge across tens of billions of SKUs. It combines human–AI collaborative ontology engineering, a retrieval-then-discrimination architecture (\(S^2D\)), and self-evolving multimodal models to achieve 94.2% precision and 82.8% recall at catalog scale. The system has accumulated hundreds of billions of knowledge assets, increased item-information richness by 3.35×, and is deployed live across search, recommendation, and merchant tooling.

Problem¶

Traditional NLP pipelines (e.g., BERT-based NER) for e-commerce item understanding suffer from two structural limitations: (1) insufficient model capacity and poor generalization to emerging concepts and heterogeneous data sources, and (2) a manual-annotation bottleneck that makes quality tightly coupled to expensive human labeling. At JD's scale—tens of billions of SKUs across virtually every retail category, with continuously emerging market concepts—these systems cannot simultaneously achieve high quality, low cost, and high throughput.

Method¶

Oxygen AIIC is organized around five tightly coupled modules:

Ontology Engineering uses a human–AI collaborative loop. Experts define a top-down backbone (category hierarchy, attribute keys/values, scenario tags). An automated bottom-up pipeline then runs three stages: knowledge discovery (a domain-fine-tuned 8B SFT model extracts attribute triples \(\langle c, x, k, d_k, v \rangle\) via OpenIE and targeted NER, yielding ~4.5M latent attribute values); knowledge fusion (a domain-adapted contrastive encoder clusters synonymous expressions, using LLM-generated synonym rewrites as positives and hard near-negatives); and knowledge validation (quality and business-importance scoring with expert sign-off).

AI Item Library uses the "Semantic Search then Discrimination" (\(S^2D\)) architecture to decouple ontology knowledge from model parameters. Rather than asking the LLM/VLM to recall ontology entries from weights, the system retrieves candidate ontology entries via semantic search and then uses the model purely for discrimination (yes/no). This eliminates hallucinations caused by ontology drift and makes the system automatically extensible as the ontology grows. Throughput is further improved by computational load reduction, cache reuse, and asynchronous pipeline parallelism on Huawei Ascend NPUs (>10× throughput gain).

Item-Understanding LLMs/VLMs are trained with multi-task SFT for instruct-following and structured extraction, then maintained via incremental learning and model self-evolution to fill targeted knowledge gaps without catastrophic forgetting.

Item Tunnel provides a tiered-freshness data and service hub supporting second-, minute-, and day-level pipelines to serve the latency/throughput tradeoffs of different downstream consumers (real-time compliance, nearline search features, offline category analytics).

Key Contributions¶

Human–AI collaborative ontology engineering framework that scales to millions of ontology entries while maintaining expert-controlled quality through a discovery–fusion–validation pipeline.
\(S^2D\) ("Semantic Search then Discrimination") architecture that decouples ontology from model parameters, eliminating hallucination on evolving catalogs and enabling plug-and-play ontology extension without retraining.
Self-evolving LLM/VLM training strategy (incremental learning + model self-evolution) enabling stable, controllable quality improvement without catastrophic forgetting.
Unified item tunnel delivering tiered-freshness knowledge services across heterogeneous downstream latency and throughput requirements.
Full industrial deployment at the scale of tens of billions of SKUs on Huawei Ascend NPUs, with documented business impact metrics.

Results¶

Knowledge production: 94.2% precision / 82.8% recall across tens of billions of SKUs.
Throughput: >10× improvement over the previous system on Huawei Ascend NPUs.
Item-information richness: increased to 3.35× its prior level.
Accumulated hundreds of billions of item-knowledge assets.
Search traffic coverage: 80.4%; item-information quality issues reduced by 37%.
Merchant listing automated attribute fill rate: >80%.
Creative optimization click-through rate improvement: ~9%.
Category planning decision cycle: reduced from weeks to days.
Discovery model on candidate knowledge extraction: 91% precision / 79% recall (8B SFT model, <2% variation across base model choices).

Limitations¶

The paper does not report ablations isolating the contribution of \(S^2D\) vs. the self-evolution component; the combined system result makes it hard to attribute gains.
Hardware dependency on Huawei Ascend NPUs means throughput figures may not transfer directly to other infrastructure.
The discovery model is trained at 8B parameters with SFT; performance on truly novel concept types (outside training distribution) is reported only anecdotally.
Knowledge fusion relies on LLM-generated synonyms for positives, which may miss domain-specific synonym pairs the LLM is unaware of.
The paper is truncated; detailed results for individual modules (e.g., fusion recall, tunnel latency SLAs) are not fully presented.
Evaluation benchmarks are internal JD datasets; external reproducibility is limited.

Relevance to Vision-Language Models¶

This paper demonstrates at the largest reported production scale how VLMs can be integrated into a multi-modal item understanding pipeline—processing images alongside text titles and detail pages—to produce structured product knowledge. The \(S^2D\) architecture offers a general pattern for grounding VLM outputs against a dynamic external knowledge base rather than relying on parametric recall, directly addressing the hallucination problem that plagues VLMs in knowledge-intensive domains. The self-evolving training strategy (incremental SFT + self-evolution without catastrophic forgetting) is directly applicable to continual learning of VLMs in open-world visual recognition settings. For VLM researchers, this work provides a concrete industrial existence proof that retrieval-augmented discrimination outperforms generative recall for structured extraction at scale.