AI HS Code Recommendation Chatbot for Mongolia Customs: Based on LLM and RAG

Mongolia Customs, processing a massive volume of export and import declarations every year, faced the challenge of accurately classifying HS Code. Actual declaration data is a mix of product names, model names, and mixed-language descriptions, making classification criteria vary from staff member to staff member and often lengthening processing time. Since HS Code is directly tied to tariff assessment, accuracy was a non-negotiable requirement.

To address this, the Mongolia Customs General Administration (MCGA) introduced an AI-powered HS Code recommendation chatbot. This system combines RAG-based Hybrid Search with LLM Multi-turn inference, and runs on Rebellions' ATOM™ NPU. It analyzes product names and descriptions entered by importers and exporters to recommend HS Code candidates that meet official classification criteria, and provides a conversational interface that captures even the nuances of Mongolian-language input.

Here's how it works:

• It presents candidates by referencing similar product data and HS Code definitions.
• Based on the user's responses, it refines recommendations or requests additional information.
• Through auto-fill functionality, it reduces manual entry errors and improves consistency in administrative processing.

As a result of the deployment, the accuracy and consistency of HS Code recommendations improved, and review time for officials was also reduced. In particular, this case is significant as an example of an NPU-based system being commercially applied to actual customs administration, and on the infrastructure side, it demonstrated performance efficiency with TPS/Watt 2.7x higher than GPU. Given its excellent operational efficiency even in constrained data center and server room environments, it has been presented as a key case underpinning the Mongolian government's future AI expansion roadmap.