Inside JD’s Oxygen AIIC: An Industrial‑Scale LLM/VLM‑Powered Product Knowledge Platform for Billions of SKUs

Problem and Motivation

In JD’s ecosystem of over 7 hundred million active users and millions of merchants, high‑quality, structured product knowledge is essential for superior consumer experience, reduced management cost, and higher operational efficiency. The industry faces three persistent bottlenecks: (1) heterogeneous, rapidly emerging product concepts; (2) massive SKU volume requiring both quality and high‑throughput production; (3) diverse downstream applications (search, recommendation, governance, merchandising) demanding different latency and freshness guarantees.

Four Core Pillars of Oxygen AIIC

Ontology Engineering : builds a dynamic, high‑coverage, high‑quality knowledge base through efficient human‑machine collaboration.

AI Product Library : maps unstructured product data to the ontology, enabling scalable knowledge consumption.

Product Understanding Model : a unified multi‑task LLM/VLM that continuously self‑evolves to improve attribute extraction and reasoning.

Item Bus (Unified Data & Service Hub) : provides tiered freshness, version consistency, storage‑compute efficiency, and a standardized service interface for downstream scenarios.

Human‑in‑the‑Loop Ontology Evolution

Experts define a skeletal ontology (category → attribute key → attribute value → scene tag) for high‑value, stable concepts. Algorithms then discover incremental knowledge from massive multi‑source data, following a three‑stage pipeline: knowledge discovery, fusion, and validation. Validation uses a cascade of quality checks, importance scoring, and expert review, yielding a million‑scale ontology with a 64.5% size increase, 80.4% traffic coverage, and a 1.44× boost in average attribute dimensions.

S²D: Semantic Search‑Then‑Discrimination Architecture

Direct generation of attribute values by a generic LLM suffers from low precision (<80%) and hallucinations. The S²D architecture first retrieves candidate ontology entries via semantic search, then discriminates the correct matches. To avoid full‑scale inference on billions of SKUs, three optimizations are applied:

Dual‑Compute Reduction : reuse results for identical or highly similar SKUs and filter low‑relevance attributes before inference.

Cache Reuse : share prompt prefixes across SKUs of the same SPU, reducing repeated pre‑fill costs.

Asynchronous Pipeline : run vector generation on NPU, semantic search on Faiss CPU clusters, and discrimination on separate CPUs, using fine‑grained data parallelism (L0) and cross‑chunk async streaming (L1) for maximal resource utilization.

These measures raise precision/recall to 92%/78.3%, increase average attribute count per SKU by 1.5×, expand knowledge assets 3.35×, and improve overall throughput by more than tenfold.

Stable, Controllable Model Self‑Evolution

The product‑understanding model rests on a unified multi‑task base augmented with two key mechanisms:

LoRA‑Based Expert Training : an improved magnitude‑principle initialization accelerates convergence and raises the capacity ceiling without extra memory overhead.

GROLE Dynamic Routing : a learned selector assigns non‑negative, sum‑to‑one weights to a pool of lightweight experts, enabling rapid domain adaptation while preserving the base’s general semantics.

Incremental experts are added for new domains without retraining the base; periodic full‑model fine‑tuning consolidates accumulated expertise. The data‑closed‑loop (badcase detection → cause analysis → synthetic repair data → value‑based selection using TANDEM and BLADE) ensures that model updates target concrete failure modes, mitigating catastrophic forgetting.

Deployment and Business Impact

The Item Bus delivers tiered freshness (offline baseline, minute‑level near‑line, real‑time high‑value streams), version‑ordered writes, and unified service APIs. It supports search, advertising, C‑end guidance, B‑end merchandising, and platform governance. Reported outcomes include:

Search quality issue rate ↓ 37%.

Core attribute auto‑fill rate > 80%.

Click‑through uplift ≈ 9%.

Precision = 94.2%, Recall = 82.8% (↑ 2.2 pp and 4.5 pp over previous stage).

Ontology construction cycle reduced from >30 days to 2 weeks.

Throughput gain > 10× thanks to cache reuse and async pipeline.

Future Directions

Next steps focus on three fronts: (1) advancing the ontology from concept‑centric to relation‑centric graphs to enable GraphRAG and agent‑based recommendation; (2) making the AI product library feedback‑driven so consumption directly triggers data‑flywheel production; (3) further reducing catastrophic forgetting via Mixture‑of‑Experts, task simplification, and knowledge‑injection prompts.