Official platform of JD Retail Technology, delivering insightful R&D news and a deep look into the lives and work of technologists.
The article presents how operations‑optimization and simulation algorithms empower supply‑chain planning—covering network design, inventory layout, and large‑scale simulation—to achieve cost reduction, efficiency gains, and enhanced user experience through advanced algorithmic solutions.
JD Retail’s engineering team tackles hard AI problems by replacing a monolithic reward model with specialized small models for ad‑image generation, deploying an LLM‑driven query‑expansion pipeline that lifts conversion rates, and pruning text‑to‑image transformers using FFT and RDP to boost throughput 40% without loss, while building comprehensive evaluation tools and a semantic smart‑assistant.
The Taro‑on‑Harmony solution lets developers write a single React‑based codebase that compiles to native HarmonyOS apps with C‑API‑driven rendering, offering both single‑thread and multi‑thread architectures, full component and CSS support, high‑performance UI, and upcoming tooling, as demonstrated by JD.com’s S‑level certified shopping app.
The paper identifies a “sandglass” bottleneck in Residual Quantization Semantic Identifiers, where middle‑layer tokens dominate, causing sparse paths and long‑tail distributions that hurt e‑commerce search performance, and demonstrates that adaptive pruning of these tokens restores accuracy and efficiency better than removing the layer entirely.
JD’s advertising platform replaces rule‑based recall with a generative large‑model pipeline that unifies e‑commerce knowledge, multimodal user intent, and semantic IDs across recall, coarse‑ranking, fine‑ranking and creative optimization, while meeting sub‑100 ms latency and sub‑¥1‑per‑million‑token cost through quantization, parallelism, caching, and joint generative‑discriminative inference, delivering double‑digit performance gains and paving the way for domain‑specific foundation models.
This work presents an end‑to‑end AI pipeline that transforms existing monocular 2D videos into immersive 3D spatial streams by combining DINO‑v2‑based depth estimation, multi‑branch view synthesis, and MV‑HEVC encoding, achieving up to 33 % BD‑Rate reduction, 31 % speed gains, state‑of‑the‑art visual quality, and real‑time production suitability, validated on the new StereoV1K benchmark and deployed in JD.Vision’s e‑commerce catalog.
JD’s supply‑chain algorithm team introduces TimeHF, a billion‑scale time‑series large model that leverages RLHF to improve demand forecasting accuracy by over 10%, detailing dataset creation, the PCTLM architecture, a custom RL framework (TPO), and superior benchmark results.
ClickHouse is an open‑source, massively parallel, column‑oriented OLAP database that integrates its own columnar storage, vectorized batch processing, pre‑sorted data, diverse table engines, extensive data types, sharding with replication, sparse primary‑key and skip indexes, and a multithreaded query engine, delivering high‑throughput real‑time analytics on massive datasets.
The paper introduces One4All, a scalable multi‑task generative recommendation framework for CPS advertising that combines few‑shot intent prompting, a Rewards‑in‑Context multi‑objective optimization, and an online model‑selection strategy, delivering 2‑3× offline HitRate/NDCG gains and notable online CTR, CVR, and commission improvements.
In 2024 the advertising team deployed an end‑to‑end AIGC pipeline that automatically creates high‑quality ad images, uses the multimodal Reliable Feedback Network and the million‑size RF1M dataset to filter outputs, builds rich offline and online multimodal representations with contrastive and list‑wise learning, and optimizes ranking architecture to deliver scalable, personalized creative selection.
