Intelligent Parcel Identification Using Large Language Models in JD Express Logistics

In modern logistics, accurate parcel information and processing efficiency are critical; JD Express faces diverse item categories and personalized customer demands, leading to high breakage and claim costs, especially for high‑value 3C and fresh goods.

The traditional manual approach yields low efficiency and error‑prone results, prompting the exploration of AI‑driven intelligent parcel recognition to enhance accuracy, speed, and overall service quality.

Analysis of 2023 data shows low matching rates caused by incomplete category libraries, varied user inputs, and B‑side merchant naming noise, which negatively impact fresh‑goods handling, value‑added service recommendations, and data availability for business decisions.

To address these issues, JD Express evaluates large models (post‑BERT/GPT era) based on performance, latency, resource consumption, integration ease, scalability, and security, ultimately selecting a model that balances accuracy and cost.

The solution integrates NLP to parse parcel descriptions, extract concise category and name fields, and match them against a three‑level category hierarchy, with separate architectures for B‑to‑B (TOB) and B‑to‑C (TOC) scenarios.

For TOB, the workflow includes: (1) sending raw item descriptions to a category‑recommendation server; (2) preprocessing the text into a prompt such as {"task":"extract category and name","description":"#宁夏农垦滩羊无脊羊排5斤#"} ; (3) invoking the large model to output {"category":"肉","name":"羊排"} ; (4) matching the result to the existing category tree (e.g., "生鲜‑肉类‑羊排"); and (5) returning the final category to the client.

For TOC, the system first attempts a full‑match lookup; if unsuccessful, it follows the same preprocessing‑model‑matching steps, yielding results like {"category":"水果","name":"苹果"} and mapping to "生鲜‑水果‑苹果".

Accuracy validation on 1,000 online orders achieved over 88% matching and >95% correctness, confirming model suitability for production.

To control costs and latency, a warm‑up cache stores previously recognized item names, allowing instant retrieval for repeated queries and drastically reducing model calls.

Post‑deployment benefits include reduced breakage and claims for fresh and fragile items, more precise value‑added service recommendations, automated category expansion, lower manual correction effort, and enriched merchant profiling for personalized offers.

Future directions involve leveraging the model for demand forecasting, route optimization, and deeper user‑profile generation, further enhancing logistics efficiency and competitiveness.