AI4Optics: private, locally hosted AI chatbot for optical metrology that leverages AI advantages while protecting company know-how

Abstract

This project “AI4Optics” addresses the challenge of knowledge preservation and utilization in small and mediumsized enterprises (SMEs) within the optical measurement industry. SMEs face the risk of losing valuable domain knowledge as experienced employees retire and industry requirements evolve. The project therefore aims to develop a domain-specific chatbot that enables efficient and secure access to internal company knowledge. The system is based on a Retrieval-Augmented Generation (RAG) architecture, combining semantic search with generative language models. User queries are matched with relevant content from a locally hosted knowledge base, and the retrieved information is used to generate context-aware responses. Company documents—such as ERP data, technical manuals, and engineering records—are processed in a structured pipeline: segmented into smaller text units, converted into semantic vectors using embedding models, and stored in a vector database for efficient retrieval. This enables the chatbot to identify relevant information even when queries differ linguistically from the original sources. A key aspect of the approach is the use of a fully local data infrastructure. The chatbot can be deployed on-premise, ensuring that sensitive company knowledge remains on internal servers and is not transmitted to external cloud services. In addition to internal documentation, this concept can be extended to external but controlled datasets, such as conference proceedings—for example, those from IMEKO. Instead of publishing such content openly on the internet, it can be stored on a secure local server and made accessible through a dedicated chatbot interface. This would allow engineers and researchers to query the complete body of conference knowledge efficiently while maintaining full control over data access. Similarly, this system can be fully customised for a specific company and utilise its own company-specific database. It is therefore a completely independent system and can be adapted for any type of business. This approach addresses a growing concern: currently, many conference proceedings are published freely online, making them accessible not only to the research community but also to external AI systems. As a result, proprietary knowledge, innovative ideas, and engineering solutions may be unintentionally used to train third-party models, potentially leading to a loss of competitive advantage. By contrast, a locally hosted chatbot ensures that knowledge remains protected within a defined environment—accessible to authorized users, but not distributed publicly or absorbed into external AI ecosystems. Initial evaluations show that such chatbot systems can effectively support engineering-related tasks, including material selection, design decisions, and troubleshooting. They provide features such as natural language interaction, transparent source referencing, multilingual capabilities, and real-time response generation. Overall, the project demonstrates the strong potential of domain-specific, locally deployed AI systems to preserve critical knowledge, enhance process efficiency, and protect intellectual property, while enabling controlled and efficient access to valuable technical information.