AssetHub AI assistant: LLM-based querying for industrial assets

Over the last decade, within the Industry 4.0 landscape, Artificial Intelligence (AI) has consolidated itself as a fundamental tool across various operational areas. More recently, the Generative AI (GenAI) paradigm has enabled the emergence of new applications, particularly the use of Large Language Models (LLMs) as natural language query interfaces for heterogeneous industrial data sources. Considering that industrial asset information is vital for decision-making at all management levels, agile and accurate access to this data is strategic. Then, in this context, this paper proposes the AssetHub AI Assistant, which is a solution based on LLMs designed to interpret and process queries over industrial assets structured based on the Asset Administration Shell (AAS) model. The methodology employs a Retrieval-Augmented Generation (RAG) architecture acting as a semantic layer, combining vector databases for context retrieval for unstructured AAS data storage. The solution validation was conducted through a benchmark comprising 120 questions categorized into easy, medium, and hard difficulty levels, covering tasks ranging from direct lookups to complex aggregations. To validate the effectiveness of the contextual enrichment, experiments included a comparative analysis against a baseline solution. Results demonstrated the robustness of the proposed approach, achieving an overall accuracy of 90.83% in retrieving the requested information.