ProRail, the Dutch government agency managing railway infrastructure, has partnered with consulting firm Valcon to develop an LLM-powered chatbot. This initiative is part of ProRail’s effort to assess how large language models (LLMs) perform in practical, real-world scenarios through a series of rapid prototyping projects.
One key challenge ProRail faces is managing the extensive set of design regulations for railway stations contained in the Rail Infra Catalogue (RIC). This database includes rules on platform widths, elevator access, and bicycle parking, accessed by contractors, architects, and inspectors. However, the existing search functionality is basic, making it difficult to locate specific information quickly.
Introducing RICO: LLM-Powered Chatbot for the Rail Infra Catalogue
Valcon developed a prototype chatbot named RICO in just one week. RICO uses a Retrieval Augmented Generation (RAG) approach to answer queries about a subset of the RIC’s regulations. It works in two steps: first, it identifies the most relevant sections of the RIC, then it generates clear, precise answers based strictly on those texts.
For example, if a user asks, “What’s the minimum required platform width?” RICO locates the exact regulation in the RIC and provides a straightforward response. Unlike general-purpose chatbots, RICO is programmed to only answer questions within its defined scope and to decline unrelated queries, ensuring accuracy and relevance.
Results from the One-Week Prototype Sprint
The initial demonstration of RICO to ProRail users received positive feedback. The chatbot provided accurate answers for most questions, including complex or conditional information. Additionally, RICO references the source material directly, which helps users verify the responses themselves.
Users noted a significant improvement over the previous search interface, making it easier and faster to retrieve the needed information from the RIC.
Next Steps and Considerations
ProRail plans to build on the insights from this prototype phase to further develop RICO. Valcon identified areas for improvement, particularly in document management. Enhancing the structure and quality of the RIC documents is expected to improve the chatbot’s accuracy.
Another important factor is the inherent probabilistic nature of LLMs, meaning 100% accuracy cannot be guaranteed. Users need to be aware of this and verify answers against the original documentation. This highlights the importance of ongoing user engagement, training, and a clear change management strategy to ensure successful adoption.
For IT and development professionals interested in learning more about building AI-driven chatbots and working with LLMs, exploring specialized courses can be valuable. Resources such as Complete AI Training offer practical guidance on AI development and deployment.
Your membership also unlocks:
