Large‑Model‑Driven Evolution of E‑commerce Search and Recommendation at JD Retail

Large models have profoundly impacted e‑commerce search ("search‑push") technology, driving it toward greater intelligence and personalization while introducing challenges like product‑knowledge hallucination, complex query understanding, personalized recommendation, and privacy/security concerns.

At the AICon Global AI Conference, JD Retail's Technical Director Zhai Zhouwei presented the company's practical innovations in applying large models to e‑commerce search, aiming to inspire the community.

1. E‑commerce Industry Evolution – Over the past decade, online retail has shifted from pure shelf‑based platforms (Alibaba, JD, Pinduoduo) to a hybrid of shelf and content e‑commerce (Douyin, Kuaishou, Xiaohongshu), emphasizing lower costs, higher efficiency, and richer user experiences through technology.

1.2 Scenario Analysis – The purchase journey is divided into pre‑purchase, purchase, and post‑purchase stages, with content platforms driving demand upstream and shelf platforms handling matching and conversion downstream.

1.3 Key Problems & Technical Challenges – JD faces challenges in reducing cost, improving efficiency, and enhancing experience, which translate into GMV, UV, UCVR, and average order value optimization.

Interaction & traffic acquisition

Intent understanding

Product recall

Relevance and multimodal matching

1.4 Technical Evolution Insight – Search technology has progressed from rule‑based text retrieval, through machine‑learning (CTR/CVR, LTR), deep learning (DNN, multimodal), to the current large‑model stage, which enables conversational interaction, long‑tail coverage, and generative retrieval.

2. Large‑Model Advantages in E‑commerce – Strong language understanding/generation, extensive knowledge summarization, transfer learning, logical reasoning, and multilingual/multimodal capabilities.

2.2 Application Issues – Product‑knowledge specificity, alignment with catalog data, image understanding, personalization limits, timeliness, high training/inference cost, and security/privacy risks.

2.3 Solutions – JD proposes an AIGC framework that includes data preprocessing (grammar filtering, perplexity scoring, quality scoring, deduplication, clustering, safety filtering, balanced data ratios), continuous pre‑training with knowledge inheritance, and domain‑specific alignment.

3. Core Technologies – Instruction learning, reinforcement from search feedback, Retrieval‑Augmented Generation (knowledge‑graph RAG, user‑profile RAG, web‑search RAG), and safety mechanisms (passive detection of prompts and outputs, active SFT & RLHF). Evaluation combines general benchmarks (MMLU, CMMLU, C‑Eval, GSM8K, GAOKAO, SuperCLUE, AlignBench) with e‑commerce‑specific benchmarks and safety scoring.

4. Practical Applications – Search interaction (query guidance, cost reduction, conversion boost), intent understanding (query & product parsing, SKU‑to‑query matching), copy generation (title, marketing copy, selling points) using multimodal alignment, relevance improvement (siamese vs. interactive models, long‑tail and long‑context handling), and next‑generation AI search featuring conversational agents and AI‑driven order fulfillment.

The article concludes that the future of e‑commerce search lies in fully AI‑driven, multimodal virtual assistants that can understand user intent, recommend optimal products, and autonomously complete transactions.