Intelligent Parcel Identification in JD Express Logistics Using Large Language Models

In modern e‑commerce logistics, accurate parcel information is crucial; JD Express faces diverse item categories and personalized demands, leading to low matching rates and high damage/claim costs.

The paper analyzes the causes of low matching rates, including incomplete category libraries, varied user inputs, and noisy B‑side product names, and discusses the impact on operations, especially for high‑value and perishable goods.

It proposes using large language models (LLMs) for intelligent parcel recognition, outlining selection criteria such as performance, latency, resource consumption, integration ease, scalability, and security.

Two architecture designs are presented: a B‑side flow where merchant‑provided descriptions are pre‑processed and sent to an LLM for category extraction, and a C‑side flow that first attempts exact matching before falling back to the LLM.

Accuracy is validated on 1,000 online orders, achieving 88 % matching and over 95 % correctness, with an operational backend allowing manual correction and category enrichment.

Cost‑effective deployment is achieved through a warm‑up cache that reuses previous LLM results for identical item names, reducing inference calls.

Future directions include richer user profiling, personalized value‑added services, demand forecasting, route optimization, and further cost reductions.

References to related research on NLP, AI, and logistics are provided.