How JD Engineers Leverage LLMs and Sparse Models to Boost Search and Ads

01. LLM‑Driven Query Expansion

JD’s post‑95 engineer "阿菌" identified that traditional neural machine translation models struggled with vague or colloquial search queries such as “养生神器,” leading to poor product recall. By redefining query expansion as a text‑generation task, her team built a three‑stage training framework that leverages a large language model’s semantic understanding to generate effective expansion keywords, dramatically improving recall and conversion rates.

02. Sparse Large Model for Advertising

Doctor‑turned‑ad‑tech lead "何言" tackled the plateau of traditional dense models by introducing sparse large models. Through systematic experiments varying data scale, dense‑parameter dimensions, and sparse‑parameter ratios, his team discovered a scaling law that balances model accuracy (AUC, GAUC) with strict latency constraints, providing a clear roadmap for model upgrades in high‑throughput ad systems.

03. Proactive Risk‑Prevention Platform

Engineer "柚子" transitioned from a vendor platform to the JD Jinli B‑side platform, confronting complex system integration challenges. He instituted a comprehensive pre‑emptive risk framework—including design reviews, pressure‑test coverage, gray‑release validation, and fine‑grained monitoring—shifting the team’s focus from reactive firefighting to proactive resilience, thereby enhancing overall system robustness.

Collectively, these three initiatives demonstrate how JD’s engineers combine cutting‑edge AI techniques, rigorous experimentation, and systematic engineering practices to solve real‑world business problems, delivering measurable improvements in search relevance, advertising efficiency, and platform stability.