How JD Advertising Leverages AI Agents to Boost Ad Operations

Introduction

Since the release of ChatGPT at the end of 2022, large models have become a hot topic. The industry believes that agents are a key direction for applying large models in daily work. This article shares JD Advertising’s practice and experience with agents.

Agent in JD Advertising

What is an Agent

Agent is a system driven by a large language model that can autonomously understand, plan, remember, and use tools to automate complex tasks.

Example: Meeting Room Booking

Traditional workflow requires opening software, selecting a room, etc. With an agent, a user can say “I am on the 12th floor, find a meeting room for 15‑16 o’clock, as close as possible.” The agent uses tools to query available rooms and book them, demonstrating the advantage over a plain LLM.

Agent in Advertising Scenarios

Service Efficiency

JD Advertising platform serves hundreds of thousands of merchants daily. Manual customer service is slow and costly. An AI agent handles routine queries, forwarding only unresolved issues to human staff, improving response speed and reducing cost.

Monitoring (盯盘) Efficiency

Agents provide smart commands such as “query last month’s data for each product line,” allowing users to retrieve data with a single step instead of navigating multiple interfaces.

Technical Implementation

RAG Capability

The RAG module consists of offline knowledge construction and an online inference engine.

Offline Knowledge Construction

Business teams prepare documents (Markdown, Excel, etc.).

Content is split into chunks.

Embedding model converts chunks to vectors.

Vectors are stored in JD Vearch vector store.

Embedding dimensions vary (e.g., 1024, 1536). Proper dimension selection is required before creating the vector table.

Online Inference Engine

When a user query arrives, it is embedded, Vearch is called to retrieve relevant knowledge, and the retrieved text is combined into a prompt for the LLM.

{"docs":[{"_id":"8684356582637079690","_score":0.7526149153709412,"_source":{"content":"搜索快车-搜索快车投放要素-定向设置（关键词定向、商品定向）-关键词定向-功能入口 ..."}},{"_id":"-1119433630625640889","_score":0.7362563610076904,"_source":{"content":"搜索快车-搜索快车投放要素-定向设置（关键词定向、商品定向）-关键词定向-关键词投放策略-关键词批量添加功能（推广管理模块）..."}}...]}

The prompt is assembled and sent to the LLM, which returns an answer.

Function Call Capability

Function calls enable smart commands, e.g., querying plans with ROI > 10. The model returns a function name and arguments, the backend executes the function, and the result is either returned directly, passed back to the model for further reasoning, or rendered with a front‑end component.

{"type":"function","function":{"name":"query_xxxx_data","description":"根据ROI、时间查询广告计划数据","parameters":{"type":"object","properties":{"startTime":{"type":"string","description":"开始时间，格式yyyy-mm-dd"},"endTime":{"type":"string","description":"结束时间，格式yyyy-mm-dd"},"roi":{"type":"string","description":"ROI查询条件，如 \"ge|3\""},"cost":{"type":"string","description":"消耗查询条件，如 \"lt|500\""}}}}}

Agent Engine and Workflow

The engine interacts with the model, orchestrates tool calls, and executes code. A custom workflow scheduler runs nodes such as knowledge recall, prompt generation, model inference, tool invocation, and result rendering. The design moved away from heavy LangChain usage to lightweight native Python implementations.

Platform and Productivity Gains

The Agent platform provides a visual designer, engine, and front‑end capabilities, enabling rapid construction of agents for various business scenarios, including customer‑complaint handling and code review automation.

Customer‑complaint Agent extracts issue content, automatically creates work tickets, and can resolve known issues via the knowledge base, reducing manual effort.

Code‑review Agent integrates with Git webhooks, uses the Agent to analyze diffs, and posts review comments automatically, improving efficiency.

Future Outlook

Large models will lower industry barriers, making AI‑assisted development accessible to non‑experts. JD plans to further enrich the Agent platform with model fine‑tuning tools and more reusable methodologies.