How JD Retail’s AI Innovations Are Transforming Supply Chains, Advertising, and E‑Commerce

In the past year JD Retail’s technology team focused on open ecosystem construction, low‑price mindset, and rapid response to business needs, delivering numerous AI‑driven innovations.

Supply Chain Innovation

JD’s intelligent supply‑chain team introduced end‑to‑end inventory management based on deep neural networks, merging prediction and optimization into a single step, which reduced cumulative error and improved decision accuracy. They also pioneered explainable AI for demand forecasting, turning the model into a white‑box that provides clear attribution for each recommendation. These technologies powered an automatic replenishment system with over 85% automation, raising the average in‑stock rate above 95% and cutting inventory turnover to roughly 30 days.

User‑Item Matching in the Post‑Behavior Sequence Era

The advertising R&D team applied privacy‑preserving pre‑training to user‑item understanding, using secure multi‑party computation and group modeling to overcome data scarcity. They introduced a sequence‑summary technique that moves the ranking model’s interaction layer ahead of the sorting stage, extending behavior sequences to ten‑thousand‑length and enriching user interest modeling.

They also built a massive sorting model (hundreds of billions of parameters) and an incremental learning framework that updates the model within minutes, dramatically improving online learning of user behavior changes.

Self‑Developed ReAct/SFT/RAG Large Model Framework

In 2023 JD launched a comprehensive LLM application framework that combines the ReAct paradigm, instruction fine‑tuning (SFT), and retrieval‑augmented generation (RAG). The framework enables large models to learn domain knowledge, improve decision‑making accuracy, and be efficiently fine‑tuned, deployed, and applied across business scenarios.

Optimizations such as compiler, operator, network, and I/O improvements boosted training performance by over 40% and supported fine‑tuning of models larger than 70 B parameters. An embedding‑based lossless compression layer and integration with the Vearch vector database accelerated information retrieval.

Search & Recommendation System Upgrade

The search‑recommendation team unified the guide‑path and algorithm‑matching technologies, introducing a pre‑training‑then‑fine‑tuning paradigm for session‑level user behavior prediction. This markedly improved traffic distribution accuracy. They also enhanced image‑search by treating it as a massive Re‑ID task, employing multi‑loss joint supervision and distributed model parallelism to increase GPU memory efficiency and training speed.

Merchant System Deep Refactor

To accommodate millions of new merchants, JD integrated native, Flutter, and Taro technologies for seamless cross‑platform functionality, applied OCR, RPA, and LLM‑based semantic understanding for automated onboarding, and leveraged multimodal models for product information generation. The resulting system reduced merchant onboarding time to about 4 minutes for individual sellers and enabled rapid, low‑cost store setup.

AIGC Technology Application

JD’s AIGC platform automates the creation of high‑quality advertising creatives. It generates diverse background images via a category generator, personalizes styles with a personalized generator, produces marketing copy using fine‑tuned large language models combined with e‑commerce data, and assembles image‑text layouts through a Plan‑and‑Render framework (PlanNet + RenderNet).

Edge Intelligence

JD built a highly quantized on‑device inference engine (≈1.9 MB) that runs across Android, iOS, and HarmonyOS, supporting multiple chip types. The engine processes billions of inferences daily with >99% success rate and is integrated into the Edge‑Intelligence SDK for recommendation, search re‑ranking, and risk control.

Data Security House

The “Security House” system implements data sandboxing, federated learning, and multi‑party secure computation, enabling data to be “usable but invisible.” It integrates with JD’s big‑data, algorithm, and security platforms, providing fine‑grained access control, hardware‑based TEE encryption, and seamless data flow across analytics pipelines.

Data Asset Upgrade

JD introduced proactive metadata‑driven materialization and dynamic cost‑aware storage decisions, achieving optimal compute‑storage allocation in real time. This approach improves query latency, reduces storage costs, and supports intelligent data diagnostics and AIGC‑enhanced query answering (chatBI).

Macro System

The “Macro” system provides LBS‑based grid operations for instant retail, integrating B2C and O2O data to match supply and demand at the city‑grid level, thereby enhancing multi‑channel efficiency and driving value growth for brands.

Low‑Cost High‑Quality 3D Modeling

JD’s end‑to‑end 3D pipeline combines a dedicated capture app, a space‑coding algorithm with a proprietary .jdv format (compressing 400 MB raw data to <10 MB), and a real‑time decoder for interactive product viewing. The solution achieves photo‑realistic quality (PSNR > 40 dB) and is already deployed on hundreds of SKUs.

These technologies, refined over 2023, have dramatically accelerated JD Retail’s digital transformation, improving efficiency, reducing costs, and delivering superior experiences for merchants and consumers alike.