How TYPO3 Content Finds Its Way into AI
AI chatbots often know only what’s in their training data, and when they lack knowledge, they cheerfully make up facts on the spot. Instead of spending a lot of time training your own AI models, Martin Helmich recommends a more pragmatic approach in his presentation at the TYPO3 Developer Days 2026: Retrieval-Augmented Generation, or RAG for short. This allows existing TYPO3 content to be integrated into AI responses at runtime—in real time and without costly retraining.
Training your own AI model sounds easier than it is. Developing cutting-edge models is hardly realistic without a budget in the millions. Fine-tuning existing models to your own content also has its pitfalls: Martin reported on an internal experiment in which a model trained on support tickets cheerfully produced made-up answers—a classic case of overfitting. On top of that, there are ongoing costs for training and infrastructure.
The much more practical alternative is called retrieval-augmented generation. In this approach, relevant content is retrieved from a database at the time of the query and provided to the language model as additional context. This is based on embeddings: content is converted into high-dimensional vectors that capture its meaning. This allows a search to find relevant results even when users do not use the exact terms.
For TYPO3, Helmich demonstrated how the Solr extension (version 14 and up) can be combined with vector search. Solr automatically handles the generation of embeddings, including rights management and multilingual support. As an alternative, he introduced the SEAL project, which is built on Symfony AI.
For the chat interface, Martin demonstrated two approaches: a quick solution using the n8n orchestration tool, and a custom-built version using Symfony AI, in which a search step and a re-ranking step improve the quality of the results. Re-ranking ensures that truly relevant content appears at the top, rather than just superficially matching results.
In conclusion, Martin advised regularly testing the quality of such systems, for example, using predefined reference answers or through “LLM-as-a-Judge” procedures, in which one AI model evaluates the responses of another.
Martin Helmich
This page contains automatically translated content.