Yeah this is something we also experimented with. We realized we needed to configure the LLM for very specific use cases for it to work as expected