Scholarlink

In recent years, the adoption of artificial intelligence models based on Large Language Models (LLMs) has revolutionised human-machine interaction in several sectors. However, the use of such models in critical industrial scenarios presents challenges related to security, privacy and the need to operate in environments with limited connectivity. This paper demonstrates the design and realisation of an avatar (or AvaTeacher) that can support the user in the maintenance of complex systems, combining advanced Retrieval Augmented Generation (RAG) techniques and LLM models optimised to operate offline. The developed system provides access to a knowledge base with the integration of Technical Manuals related to an industrial system developed and designed by Leonardo S.p.A., providing accurate and contextualised answers without the need to connect to external cloud services. Moreover, one of the key features of the system is the possibility to update knowledge and add new systems without the need to train the model, thanks to the use of a modular architecture and advanced information retrieval techniques. The implementation includes the integration of advanced indexing and information retrieval strategies, ensuring high system efficiency and reliability. This flexibility allows the system to be adapted to new needs and operational scenarios with minimal intervention, reducing costs and maintenance time. Furthermore, the advantages and limitations of the adopted approach were analysed and compared with existing solutions. The results demonstrate the feasibility and effectiveness of an offline virtual assistant to support maintenance operations and training in industrial environments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Avateacher: Enhancing Industrial Maintenance with Private-Network LLMs and Retrieval-Augmented Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Salvatore Musto,
  • Nazaro Aversano,
  • Francesco de Pandi,
  • Luciano Rosario Grimaldi,
  • Salvatore D’Onofrio,
  • Francesco Bonavolontà,
  • Mauro Galateo,
  • Claudio Dotani,
  • Flora Amato

摘要

In recent years, the adoption of artificial intelligence models based on Large Language Models (LLMs) has revolutionised human-machine interaction in several sectors. However, the use of such models in critical industrial scenarios presents challenges related to security, privacy and the need to operate in environments with limited connectivity. This paper demonstrates the design and realisation of an avatar (or AvaTeacher) that can support the user in the maintenance of complex systems, combining advanced Retrieval Augmented Generation (RAG) techniques and LLM models optimised to operate offline. The developed system provides access to a knowledge base with the integration of Technical Manuals related to an industrial system developed and designed by Leonardo S.p.A., providing accurate and contextualised answers without the need to connect to external cloud services. Moreover, one of the key features of the system is the possibility to update knowledge and add new systems without the need to train the model, thanks to the use of a modular architecture and advanced information retrieval techniques. The implementation includes the integration of advanced indexing and information retrieval strategies, ensuring high system efficiency and reliability. This flexibility allows the system to be adapted to new needs and operational scenarios with minimal intervention, reducing costs and maintenance time. Furthermore, the advantages and limitations of the adopted approach were analysed and compared with existing solutions. The results demonstrate the feasibility and effectiveness of an offline virtual assistant to support maintenance operations and training in industrial environments.