AI Agent Applications and Architecture in the 1688 E‑commerce Platform

The presentation introduces the new capabilities that large language models (LLMs) bring to e‑commerce, outlining six fundamental abilities—generation, summarization, extraction, rewriting, classification, and retrieval—that underpin AI agents for solving diverse business problems on the 1688 platform.

It describes how AI is integrated into the e‑commerce workflow, enhancing stages such as product discovery, recommendation, inquiry handling, and backend operations, and proposes a unified data‑to‑decision pipeline that transforms user intent into intelligent actions.

The speaker details 1688’s specific AI agent exploration, highlighting the use of the Tongyi Qianwen‑72B model with a 32K context window, its performance advantages over smaller models, and the transition from plan‑based to auto‑agent modes for more autonomous task execution.

A comprehensive agent solution is presented, including LLM‑driven application patterns, prompt engineering, a dual‑gateway routing architecture, knowledge‑base vector retrieval, and a tree‑structured task execution engine that supports both single‑turn and multi‑turn interactions.

Deployment considerations focus on GPU‑centric inference, resource scheduling, caching strategies, and scalability challenges unique to large models, as well as quality assurance processes that combine automated metrics, human evaluation, and GPT‑4‑based assessments.

The talk concludes with an AI innovation paradigm shift, emphasizing product‑driven development, agile R&D workflows, and the need for robust engineering support to accelerate model fine‑tuning, prompt stability, and platform integration.

Future plans include expanding AI capabilities across consumer and industrial e‑commerce, improving supply‑chain automation, and exploring video and digital‑human applications to enhance user experience.

A Q&A session addresses differences between agents and RPA, strategies to avoid over‑fitting during fine‑tuning, and the trade‑offs between traditional search/recommendation systems and LLM‑enhanced solutions.