PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Oxygen AIIC, 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

arXiv:2606.28070v2cs.AI

TL;DR

Large-scale e-commerce requires item knowledge that addresses incomplete information, semantic gaps, massive SKU volume, and diverse downstream needs. Oxygen AIIC combines ontology engineering, LLM/VLM-based knowledge production, model evolution, and unified serving, achieving 94.2% precision and 82.8% recall at industrial scale.

  • Problem

    Large-scale e-commerce item-knowledge systems face incomplete information, semantic gaps, and industrial-scale demands across massive catalogs and downstream applications.

  • Method

    Oxygen AIIC integrates ontology engineering, LLM/VLM-based knowledge production and evolution, modular infrastructure, and a unified item tunnel for serving applications.

  • Results

    94.2% precision and 82.8% recall were achieved for knowledge production across tens of thousands of categories and tens of billions of SKUs.

  • Takeaways & Limitations

    Oxygen AIIC validates the industrial feasibility of LLMs/VLMs when coordinated with ontology engineering, engineering systems, and business feedback loops.

  • Takeaways & Limitations

    Relation modeling remains preliminary and has not yet reached scale, limiting the ontology’s potential for applications and efficient industry-knowledge use.

Abstract

from arXiv · show

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%.

1 Introduction

JD.com’s massive, rapidly changing catalog exposes demand-, supply-, and operations-side limits in traditional item-knowledge systems. Oxygen AIIC addresses the resulting industrial deployment challenges with an LLM/VLM-centered platform for evolving ontologies, scalable knowledge production, and diverse downstream services.

  • JD.com serves over 700 million active users and millions of merchants while managing tens of billions of SKUs.
  • Industrial-scale bottlenecks: Traditional systems face incomplete information and semantic gaps, costly manual item management, and fast-changing trends requiring finer-grained operations.These bottlenecks respectively degrade user experience and traffic allocation, lower product-information quality, and constrain platform efficiency.
  • Deployment challenges: JD’s industrial deployment must evolve ontologies for emerging concepts, produce high-quality knowledge cheaply at massive scale, and serve scenario-specific requirements efficiently.The downstream requirements span real-time governance, high-throughput nearline search and recommendation, and offline category operations.
  • Oxygen AIIC: Oxygen AIIC constructs an ontology with millions of entries and produces item knowledge across tens of thousands of categories and tens of billions of SKUs.Its architecture combines human–AI ontology expansion, S2D knowledge production, self-evolving item-understanding LLMs/VLMs, and a unified item tunnel.
  • Oxygen AIIC: 94.2%/82.8% knowledge-production precision/recall and a more than 10× throughput-efficiency gain demonstrate Oxygen AIIC’s industrial-scale performance on Huawei Ascend NPUs.
  • Oxygen AIIC: The unified item tunnel acts as a shared data and service hub, maintaining freshness through tiered service levels while supporting diverse downstream applications.Together with the application matrix, it supports a wide range of business scenarios and forms a sustainable business ecosystem.

2 Architecture Overview

Oxygen AIIC uses five tightly coordinated modules in a modular, low-coupling architecture that supports scalable item-knowledge production, centralized serving, and continuous evolution. Together, these modules form an end-to-end loop connecting ontology construction, knowledge production, model improvement, unified access, and business applications.

  • Architecture: Five tightly coordinated modules—ontology engineering, AI Item Library, item-understanding LLMs/VLMs, item tunnel, and applications—form Oxygen AIIC’s modular architecture.The modules are decoupled so core capabilities can be iterated independently, improving development efficiency, maintainability, and stable evolution.
  • Ontology Engineering: Ontology engineering combines JD expert knowledge with LLM/VLM world knowledge and reasoning through human–AI collaboration to build the platform’s knowledge foundation.Experts distill industry knowledge while algorithms learn from it to scale ontology construction.
  • AI Item Library: The AI Item Library maps tens of billions of items to an evolving ontology through a jointly optimized model-data-engineering pipeline for scalable, high-throughput knowledge production.Separating ontology from model parameters reduces hallucinations and improves generalization, while computational-load reduction, cache reuse, and asynchronous processing improve throughput.
  • Item-Understanding LLMs/VLMs: Item-understanding LLMs/VLMs support ontology construction and AI Item Library production, using incremental learning and self-evolution to fill knowledge gaps and mitigate catastrophic forgetting.The models provide a generalizable and scalable foundation for continuous improvement in data quality.
  • Item Tunnel and Applications: The item tunnel provides a unified service layer with daily-, minute-, and second-level production and distribution pipelines, while applications deploy standardized capabilities across business scenarios and the item lifecycle.The architecture’s integrated modules create a closed loop spanning ontology construction, knowledge production, asset management, tiered access, and cross-domain feedback.

3 Ontology Engineering

Oxygen AIIC uses expert-defined ontology standards with LLM/VLM-driven discovery, fusion, and validation to evolve a broad, high-quality ontology continuously. This human–AI collaboration addresses emerging concepts, semantic redundancy, and the tension between scalable coverage and controllable quality.

  • Human–AI collaboration: Experts define ontology standards and audit quality, while LLMs perform large-scale discovery, expansion, and refinement under expert guidance.The framework enables continuous ontology evolution while maintaining high quality and broad coverage.
  • Ontology structure: The ontology backbone organizes downstream requirements through four element types: Category, Attribute Key, Attribute Value, and Scenario Tag.These elements represent taxonomy, feature dimensions, specific attribute instantiations, and higher-level consumption concepts.
  • Discovery–fusion–validation: The bottom-up pipeline discovers emerging concepts from heterogeneous sources, fuses synonymous candidates, and validates their quality and business importance.Experts establish the backbone, while algorithms incorporate user-behavior and industry-trend signals at scale.
  • Knowledge discovery: 91% precision and 79% recall are achieved on candidate knowledge extraction by the dedicated e-commerce knowledge-discovery model.The model is trained for domain-specific concepts, industry standards, and emerging terminology that general-purpose models often miss.
  • Knowledge discovery and fusion: 4.5 million latent attribute values are mined from item information, user queries, and external web content, then reduced to 2.1 million candidate ontology concepts.The fusion pipeline consolidates approximately 2.4 million redundant concepts into synonym sets.
  • Scale and impact: 64.5% ontology expansion and 1.44× the previous level in average characterization dimensions per item increase product-information richness, with coverage reaching 80.4% of JD’s user traffic.The resulting ontology is million-scale and includes attributes, product terms, brands, and other entities.

4 AI Item Library

The AI Item Library maps unstructured item information to a dynamically evolving ontology at industrial scale, supporting high-freshness applications across JD’s catalog. Its central challenges are adapting to ontology evolution without costly retraining and avoiding throughput bottlenecks across massive, heterogeneous SKU data.

  • AI Item Library: The library builds scalable, extensible semantic mappings from unstructured item information to ontology elements, forming stable item-to-ontology association chains.The ontology supplies a standardized knowledge foundation, while the library performs high-precision knowledge identification.
  • AI Item Library: Tens of billions of SKUs, tens of thousands of categories, and millions of evolving ontology entries define the library’s industrial-scale data environment.It must also process hundreds of millions of daily item-information changes while maintaining high freshness for JD’s core application scenarios.
  • Core Technical Challenges: Dynamic ontology evolution challenges conventional end-to-end models because ontology knowledge is tightly coupled to parameters, causing costly adaptation and degraded out-of-distribution performance.Frequent fine-tuning or retraining to accommodate ontology updates incurs prohibitive computational and time costs.
  • Core Technical Challenges: Exhaustively evaluating all attributes for tens of billions of SKUs creates a throughput bottleneck through redundant processing of homogeneous variants and unproductive computation on sparse long-tail attributes.The challenge arises under massive data volumes where most sparse long-tail attributes provide no useful signals.

4.1 Method Overview

Oxygen AIIC addresses industrial item-knowledge recognition through a collaborative optimization system spanning models, data, and engineering. Its method combines the S2D architecture, which separates ontology updates from model parameters, with SKU- and attribute-level computational load reduction.

  • Existing item-knowledge recognition methods face significant structural limitations in real-world industrial scenarios.
  • The proposed system treats scalability and high throughput as joint design objectives across models, data, and engineering.
  • Semantic Search then Discrimination: S2D decouples the ontology from model parameters by externalizing the dynamically evolving ontology into a separate knowledge base.Semantic encoders retrieve relevant ontology entries, after which the model determines whether the retrieved entries apply.
  • Throughput Optimization: Computational load reduction targets both SKU and attribute dimensions through similarity-based deduplication and highly relevant attribute probing.Cache reuse and asynchronous pipeline parallelism further improve efficiency.

4.2 Item Knowledge Recognition

Item knowledge recognition uses a Semantic Search then Discrimination pipeline: semantic retrieval narrows ontology matches to Top-K candidates, and constrained discrimination performs precise selection. S2D achieves 92% precision and 78.3% recall while expanding item-knowledge assets to hundreds of billions.

  • Semantic Search then Discrimination: S2D retrieves Top-K ontology candidates in a shared semantic space, then applies fine-grained discrimination within the recalled candidate set.The representation model supports newly added ontology entries without model retraining, while the discrimination model restricts outputs to ontology-constrained candidates.
  • Semantic Search then Discrimination: K = 10 balances strong recall with low downstream discrimination cost and reduces the computational burden of candidate filtering.Semantic search ranks candidates by cosine similarity between item and attribute-value representations.
  • Semantic Search then Discrimination: Ontology-constrained set-based filtering improves controllability by selecting values from recalled candidates instead of freely generating attribute values.The discrimination model transfers a large-scale reference model’s expertise into a more efficient 8B-parameter model and mitigates hallucination risk.
  • Evaluation: 92% precision and 78.3% recall are achieved by S2D under the evaluation protocol.These results demonstrate high recognition accuracy for item knowledge.

4.3 Throughput Efficiency Improvement

Oxygen AIIC improves throughput by shifting from exhaustive SKU–attribute computation to differentiated SKUs and highly relevant attributes, using load reduction, cache reuse, and asynchronous pipeline parallelism. Together, these optimizations deliver more than tenfold higher overall throughput efficiency while preserving recognition accuracy.

  • Optimization objective: Three complementary strategies—computational load reduction, cache reuse, and asynchronous pipeline parallelism—shift processing toward differentiated SKUs and highly relevant attributes.The objective reduces compute costs while preserving recognition accuracy.
  • Computational load reduction: 3x throughput efficiency results from SKU-level semantic deduplication and attribute-level sparsification in the Oxygen AIIC production pipeline.SKU deduplication reuses recognition results for homogeneous variants, while sparsification probes only relevant attributes.
  • Cache reuse: 85% of item information is shared across SKUs under the same SPU, motivating shared-prefix prompts, SPU-aware routing, and tuned cache memory management.The empirically optimal block_size=16 is compared with the default value of 128.
  • Cache reuse: More than sixfold production-throughput improvement comes from prompt-structure, cache-locality, and memory-management optimizations.These optimizations maximize prefix-cache hit rates, prevent cache eviction, and limit block-management overhead.
  • Asynchronous pipeline parallelism: More than twofold overall system-throughput improvement comes from asynchronous pipeline parallelism, which reduces waiting latency across heterogeneous hardware.The architecture targets latency caused by heterogeneous hardware execution.
  • System-level result: More than tenfold overall throughput-efficiency improvement results from jointly applying computational load reduction, cache reuse, and asynchronous pipeline parallelism.The three optimizations exhibit strong scaling effects as the ontology and attribute dimension expand.

5 Item Understanding LLMs/VLMs

Oxygen AIIC introduces a unified, multi-task item-understanding LLM/VLM framework with incremental capability expansion, instruction-following knowledge representation, and closed-loop self-evolution. The resulting system improves scalable item-knowledge production while supporting continuous model iteration and stable performance.

  • Framework motivation: The framework addresses weak domain knowledge, costly retraining, noisy item features, and subtle long-tail recognition defects through unified modeling and controlled iteration.These challenges limit off-the-shelf performance, data reuse, adaptation efficiency, and defect repair.
  • Unified foundation: Unified multi-task modeling consolidates knowledge extraction and recognition tasks, giving the models generalized understanding of items, ontology entries, and task patterns.Extraction includes attribute keys, values, and key-value pairs; recognition includes key, value, and key-value discrimination.
  • Incremental capability expansion: Task-free lightweight expert modules expand capabilities for emerging ontology entries without fully retraining the foundation model or sacrificing existing capabilities.The system dynamically integrates incremental experts into an expert pool and periodically consolidates them through full fine-tuning.
  • Incremental capability expansion: LoRAM improves incremental fitting efficiency and raises the performance ceiling by optimizing adapter magnitude dynamics without additional memory overhead or preprocessing cost.Conventional LoRA’s low-rank updates can impair early fitting dynamics, slow convergence, and limit attainable accuracy.
  • Model self-evolution: A four-stage self-evolution framework—data evaluation, analysis, synthesis, and selection—forms a closed loop that targets defects instead of indiscriminate data expansion.Its outputs include defect type, impact scope, and a suggested repair strategy.
  • Evaluation results: 94.2% precision and 82.8% recall are achieved in end-to-end AI Item Library production, while only 0.8% of attributes lose over 5% precision and new-entry turnaround falls to approximately two weeks.Precision and recall increase by 2.2% and 4.5%, respectively, compared with Section 4.2.3; turnaround falls from over 30 days.

6 Item Tunnel

Oxygen AIIC’s unified item tunnel serves as the data-and-compute hub connecting ontology construction, AI Item Library production, model self-evolution, and downstream applications. It supports massive-scale updates and item coverage while balancing freshness, consistency, storage-compute efficiency, and service reuse.

  • Item Tunnel: The unified item tunnel connects knowledge production and downstream applications as Oxygen AIIC’s shared data-and-compute hub.It supports ontology construction, AI Item Library production, and model self-evolution while enabling efficient consumption.
  • Item Tunnel: Hundreds of millions of AI Item Library updates per day support a dynamic library covering tens of billions of items.The infrastructure maintains eventual consistency across heterogeneous service modalities, with convergence achieved within minutes.
  • Tiered-freshness pipelines: Freshness tiers spanning seconds, minutes, and days balance data-change frequency, business value, inference latency, throughput, NPU utilization, and compute cost.Offline inference offers higher throughput and NPU utilization but longer latency, whereas real-time inference responds quickly but can repeatedly process transient states.
  • Eventual consistency: Eventual-consistency mechanisms resolve data-version divergence across cascaded production paths and control convergence time down to the minute level.The hybrid stream-batch architecture must prevent incorrect version overwrites and data loss caused by distributed production, replay, or delayed data.
  • Storage-compute efficiency: Physical decoupling, elastic scheduling, and a unified data contract support stable production across data flows covering tens of billions of items.The design coordinates heterogeneous Spark/Flink and vLLM inference clusters through offline HDFS Parquet shards and real-time Kafka partitions.
  • Unified services: A standard service matrix replaces fragmented integrations, unifying offline storage, containerization, inference engines, and service frameworks for scalable reuse.The tunnel exposes standard service classes for online item and ontology lookup, high-throughput relational-table consumption, and event-driven full or incremental access.

7 Applications

Oxygen AIIC productizes unified item knowledge across consumer shopping, item management, and platform governance. Its applications improve search and shopping assistance, accelerate operational decisions, raise listing and creative quality, and support ecosystem governance.

  • Consumer shopping: Search item-information richness reaches 3.35× the previous level, while item-information defects in search fall by 37%.The AI Item Library and ontology also support search guidance and faceted filtering.
  • Consumer shopping: Structured item knowledge enables conversational e-commerce and AI comparison, helping users handle long-tail needs and compare candidate items efficiently.Assistants support natural-language item discovery, specification checking, scenario matching, and structured reports covering highlights, parameters, and reviews.
  • Item management: Category planning shortens decisions from two or three weeks to a few days, while automated core-attribute filling exceeds 80%.The end-to-end item-management loop spans category planning, product listing, and product operations across JD business formats.
  • Item management: Product operations increase click-through rate by about 9% by standardizing and enhancing multimodal item creatives through low-cost A/B testing.The system distills visual standards for resolution, color, composition, and information hierarchy from high-conversion images.
  • Platform governance: Oxygen AIIC supports assortment selection, audience profiling, and product-information governance through richer item data and linked user-behavior analysis.These capabilities improve selection quality, identify high-potential users, and proactively intercept or inspect non-compliant product information.
  • Platform governance: Identical-item recognition exceeds 90% accuracy across JD Retail physical categories, supporting competitive and standardized pricing decisions.Price governance combines recognition with a price-rating system for merchants and category buyers.

8 Related Work

Prior e-commerce research advances knowledge graphs, ontology expansion, item knowledge production, and domain foundation models, but largely addresses isolated stages or model capabilities. A unified infrastructure coordinating ontology evolution, scalable production, validation, serving, and feedback remains insufficiently explored.

  • E-commerce knowledge graphs: E-commerce knowledge graphs structure item relations, user intents, and commonsense knowledge for organization, recommendation, and explainable ranking.Representative approaches connect shopping intents to item organization, mine purchase intents with human-in-the-loop LLM pipelines, and scale commonsense production through critic-guided generation.
  • E-commerce knowledge graphs: These graph studies focus on individual construction or application stages, leaving coordinated production, validation, deployment, and feedback insufficiently explored.The related work improves structured representation but does not provide closed-loop coordination across the full knowledge lifecycle.
  • Ontology expansion: Ontology research adapts taxonomies to emerging item types and discovers category-attribute relations, but separates ontology evolution, schema recognition, and SKU-level completion.Octet uses heterogeneous query, item, and category relations for self-supervised taxonomy enrichment, while KATIE addresses attribute relations, importance, and synonym merging.
  • Item knowledge production: Item knowledge production has progressed from open extraction and large-scale label expansion to multi-source, multimodal modeling, and automatic validation.However, existing methods devote limited attention to production and management efficiency at massive scale, limiting industrial deployment feasibility.
  • E-commerce domain foundation models: E-commerce foundation models improve item understanding, cross-task transfer, and cold-start generalization, but primarily emphasize model capabilities rather than integrated industrial infrastructure.The missing infrastructure concerns dynamic ontology evolution, production cost, result verifiability, hybrid serving, and downstream feedback loops.

9 Conclusion

Oxygen AIIC is an industrial-scale LLM/VLM-centric infrastructure that integrates ontology construction, knowledge production, model evolution, item serving, and downstream applications. Its conclusion emphasizes that scalable e-commerce knowledge construction requires coordinated system-level capabilities rather than isolated models.

  • System contribution: Oxygen AIIC provides an end-to-end infrastructure for item knowledge production and serving across large-scale e-commerce scenarios.The system spans ontology construction, knowledge production, model evolution, the item tunnel, and downstream business applications.
  • Ontology engineering: Human–AI collaboration dynamically discovers, fuses, validates, and expands a high-quality item ontology.Expert domain knowledge is combined with large models’ generalization and reasoning capabilities.
  • Knowledge production: Tens of billions of SKUs are supported through semantic search then discrimination, ontology–model decoupling, and throughput optimizations.Computational-load reduction, cache reuse, and asynchronous pipeline parallelism enable high-throughput, low-cost knowledge production.
  • Model system: The multi-task item-understanding model uses incremental learning, instruction-following representations, and self-evolution to improve steadily and repair long-tail defects.The model system is designed to avoid systemic degradation in a constantly changing item ecosystem.
  • Industrial deployment: Tens of thousands of categories and tens of billions of SKUs are covered, with hundreds of billions of item-knowledge assets accumulated.The unified item tunnel delivers knowledge securely and reliably to search, recommendation, listing, governance, operations, merchant recruitment, assortment review, and identical-item recognition.

10 Limitations and Future Work

Despite Oxygen AIIC V1’s significant results, ultralarge-scale industrial deployment still faces fundamental challenges, including preliminary relation modeling that limits application support. Future work targets stronger ontology engineering and online bad-case discovery supported by a data flywheel and model self-evolution.

  • Limitations: Ontology relation modeling remains preliminary and has not yet reached scale, limiting Oxygen AIIC’s ability to empower applications and use JD’s accumulated knowledge.The current version has substantially expanded the ontology’s conceptual coverage, but relation modeling remains a fundamental challenge.
  • Future Work: Future ontology engineering will expand relation modeling, enrich the ontology, and enable graph-based reasoning.The plan also incorporates expert e-commerce knowledge and industry know-how for reuse across business scenarios.
  • Future Work: Future work will develop an online mechanism for bad-case discovery coupled with a data flywheel and model self-evolution.
Loading 2606.28070v2…