How SF Express Secures AI Agents for Logistics: Responsibility, Safety, and Performance

When Baishi Group boosted cross‑border document review rates by 30% with DeepSeek and Shaanxi logistics opened policy‑interpretation chains via the "QinWutong" model, SF Express emphasizes that the real battlefield for logistics large models is not flashy technology but achieving industrial‑grade closed‑loop decision making that is responsible, safe, professional, and collaborative.

Facing industry‑wide hallucination issues that can cause cargo‑plane load imbalances and scheduling errors, SF Express built an anti‑hallucination system composed of a rule‑compilation layer, logistics knowledge base, and dynamic circuit‑breaker, improving on‑time air‑dispatch rates by 7.2% over manual operations. To address data‑security concerns, they deployed 100% private‑domain servers, intercepting 120,000 abnormal accesses and reducing leakage risk by 90% compared with public‑cloud solutions.

DataFun: How do you achieve "responsibility" and "safety" in logistics intelligent agents?

Liu Ziheng: Responsibility means the agent must deliver accurate, timely, complete, and explainable results. Safety means using core business knowledge while guaranteeing information security. We ensure responsibility by clearly defining service scope and enhancing performance (reducing hallucinations, improving inference speed, optimizing outputs). Safety is achieved through strict data‑governance, private‑deployment, and monitoring mechanisms.

DataFun: What are the key milestones in the evolution of SF’s intelligent agents?

Liu Ziheng: Early stages focused on native large‑model capabilities—prompt engineering and knowledge‑base construction—for simple Q&A. Mid‑to‑high stages address complex logistics decisions by combining LLMs, tool‑calling, memory modules, and planning logic (e.g., ReAct), enabling multi‑step tasks and coordination with domain‑specific sub‑models.

DataFun: How does SF transform logistics expert experience into computable agent components?

We employ several methods:

Model fine‑tuning for domain‑specific tasks.

Orchestrating large and small models (e.g., optimization sub‑models) for flexible agents.

Building a logistics knowledge base to answer professional queries.

Prompt engineering tailored to logistics decision problems.

Integrating business workflows so the model learns decision processes.

DataFun: How do you balance automation with human intervention in agent interactions?

Automation level depends on scenario complexity and data completeness. Clear, data‑rich tasks (e.g., resource adjustments) can be highly automated, while exploratory or strategic planning still requires human oversight.

DataFun: Which capabilities must remain fully self‑controlled?

Core logistics network optimization—where SF demands industry‑leading execution efficiency—must be developed in‑house and remain autonomous.

DataFun: Does SF provide low‑code platforms for business users?

Yes, platforms like "Fengyu Assistant" enable low‑code development, and tools such as Dify and LangGraph support more flexible agent creation. Training courses, reference cases, and internal competitions help business staff quickly build and deploy agents.

DataFun: What technical bottlenecks must be overcome for large‑scale agent deployment in logistics?

Hallucination: High‑accuracy decisions demand minimal model‑generated fabrications.

Inference performance: Handling massive, real‑time data streams requires efficient context processing and low latency.

Tool‑calling ability: Agents must seamlessly interact with operational systems, data platforms, and hardware.

Focusing on specific domain problems rather than attempting a universal agent service yields more practical results.