AI at KMSKA: A New Challenge
The Royal Museum of Fine Arts Antwerp (KMSKA) has recently launched an innovative AI project, which is now in its final preparation stages. What began as a practical tool to assist staff with visitor questions has grown into a fascinating experiment at the crossroads of art, technology, and audience engagement. A glimpse behind the scenes.
No Question Too Many
The idea originated with the restorers and curators working since autumn 2023 on the Studio Rubens restoration project. They noticed that visitors often approached them with practical questions about navigation or facilities in the museum. To support the team, the idea of a chatbot soon emerged. Initially meant as an additional info point, it quickly expanded to cover more content-driven, art-historical questions. The ICT department deliberately chose to develop the project in-house, keeping all data within the museum while gaining valuable insights into the possibilities of AI.
Step by Step
After thorough research, the chatbot was built using open-source tools such as Ollama, Whisper, and Piper TTS. The concept is simple: visitors at Studio Rubens can ask a question out loud, after which the software converts the audio into text, processes it, and replies with a natural-sounding voice. Here’s how it works: Ollama serves as the server application hosting the Large Language Model (LLM). Once a visitor asks a question, Whisper transcribes the audio into text. This text is then combined with the relevant knowledge and sent to Ollama. The generated answer, initially in text, is converted into audio by Piper TTS, in Dutch or English, and sent back to the AI kiosk in Studio Rubens.
Challenges
The biggest obstacle was hardware: Large Language Models require substantial memory and computing power, while only consumer-grade equipment was available. This meant opting for a smaller model and striking a balance between speed and quality. Multilingual support also posed difficulties. While English was well recognized, initial problems arose with Dutch spelling and grammar. The most effective solution turned out to be using Dutch prompts combined with an English-Dutch mapping for accurate results.
Conclusion
With its first AI project, the KMSKA has succeeded in transforming a theoretical idea into a working proof of concept. Even though it is still in its early stages, it offered the IT-team a valuable opportunity to explore AI and promises to enhance the visitor experience in the near future. By experimenting and learning from mistakes, the museum identified what works, where the limitations lie, and how issues such as multilingualism can be improved. With these insights, KMSKA hopes to inspire others to apply technology in creative and meaningful ways.
