ProRail, the organization managing railway infrastructure in the Netherlands, has partnered with consulting firm Valcon to develop and test an innovative chatbot powered by large language models (LLMs). The initiative is part of ProRail´s broader exploration of LLMs, launching five rapid prototyping ´pressure cooker´ projects to evaluate the technology in real-world applications. In their collaboration with Valcon, the focus was placed on enhancing access to the extensive Rail Infra Catalogue (RIC), a key database of regulations for railway station design, which is frequently used by contractors, architects, and inspectors throughout the country.
Traditionally, users found it difficult to navigate the RIC´s vast content because of limited, basic search features. To address this challenge, Valcon and ProRail created ´RICO,´ an LLM-based chatbot utilizing Retrieval Augmented Generation (RAG) techniques. Within a single week, Valcon developed a prototype able to respond to detailed regulatory queries by first retrieving relevant content from the RIC and then using the LLM to generate clear, contextually appropriate answers. A key feature of RICO is its ability to provide direct references to source material and restrict responses strictly to the information available in the RIC, thereby avoiding the pitfalls of overly broad or hallucinated answers that sometimes affect general-purpose chatbots like ChatGPT.
User feedback after the one-week pilot was highly positive, with RICO demonstrating expertise in supplying accurate, readable answers to both simple and complex inquiries, and offering significant improvements over the older search interface. ProRail now plans to use the insights from this pilot to advance the chatbot from proof-of-concept to a more robust and production-ready tool. The evaluation also surfaced areas for further enhancement, including refining document management to improve accuracy and implementing more comprehensive user training and change management strategies, acknowledging that the probabilistic nature of LLMs means responses should always be validated against official documentation.
