Interview with JD Retail AI Director Zhai Zhouwei on the Evolution and Future of E‑commerce Search Powered by Large Models

As e‑commerce continues to grow, search technology becomes a crucial bridge between users and products. JD Retail’s AI Director Zhai Zhouwei discusses the evolution of e‑commerce search, dividing it into four stages: text retrieval, machine‑learning‑based intent and product understanding, deep‑learning‑driven multimodal retrieval, and the current large‑model era.

In the large‑model stage, interaction shifts to bidirectional conversational interfaces, and large models improve long‑tail generalization, recall, and relevance, including generative retrieval techniques.

Zhai explains JD’s large‑model applications: conversational assistants (e.g., Jingyan AI), enhanced intent and product understanding, augmented product recall, relevance matching, AI‑generated marketing copy, and comment summarization.

He identifies key deployment challenges: limited product knowledge in generic models, personalized context understanding, timeliness of knowledge updates, high training and inference costs, latency constraints, and security/compliance risks.

To address real‑time information needs, JD employs continuous pre‑training with fresh data and Retrieval‑Augmented Generation (RAG) pipelines such as KG‑RAG, product‑search RAG, and web‑search RAG, achieving superior performance over generic models.

Building an effective large‑model e‑commerce search engine requires a high‑performance model backbone, robust evaluation metrics, and continuous learning methods like inheritance learning, knowledge‑density tuning, model‑structure optimization, annealing, multi‑stage instruction alignment, and safety governance.

Regarding ROI, Zhai notes that while many large‑model applications are still loss‑making, they hold huge commercial potential; optimizing existing business models with large models can yield measurable ROI, especially for long‑tail problems.

He envisions the next‑generation AI e‑commerce search as a fully large‑model‑driven digital assistant capable of multimodal, natural‑language interaction, autonomous ordering, and end‑to‑end logistics via AI agents.

For enterprises, Zhai advises starting with large‑model cloud APIs for rapid prototyping, then considering open‑source, performance‑oriented models for deeper optimization, while recognizing the high resource and data costs of custom model development.

The interview concludes with an invitation to Zhai’s upcoming talk titled “E‑commerce Large Models and Search Application Practices” at AICon Shanghai on August 18‑19.