At the TYPO3 Happy Hour on 5 December 2025, I had the opportunity to present a project that is very close to my heart: The new AI chatbot for the University of Erfurt.
While the bot itself runs on a Python stack, its functionality is entirely dependent on the content managed within our TYPO3 ecosystem. In my talk, I focused on the practical implementation of a RAG (Retrieval-Augmented Generation) architecture, moving beyond the hype to demonstrate how we make over 15,000 web pages accessible via a conversational interface.
The challenge: Making information accessible
The University of Erfurt’s website is huge. For students, finding specific information, such as enrolment deadlines or course details, can sometimes feel like searching for a needle in a haystack.
In my presentation, I demonstrated how we solved this by building a custom chatbot that:
- crawls the website incrementally (saving resources);
- vectorises the content for semantic search using FAISS;
- generates accurate answers using an LLM (Mistral AI) based strictly on our data.
Key takeaway: Quality content is crucial
The most important insight that I shared with the TYPO3 community wasn’t about Python code, but about content strategy. ‘Garbage in, garbage out’ applies heavily to AI.
We learnt this the hard way when the bot initially insisted that the winter semester started in 2019, as it prioritised an old PDF over the current landing page. This emphasised the importance of clean HTML structures and up-to-date metadata in the CMS for AI integration. This also connects directly to my previous talk about structured data and schema.org, clean data makes for smarter bots.
Slides
If you missed the talk or would like to explore the code snippets related to the crawler logic, vector database or caching mechanisms we employed, the slides are now accessible online.
The presentation runs directly in your browser:
👉 View slides: Building a RAG Chatbot
Thank you to everyone who attended!
English