Towards open-domain conversation in Rasa: Blenderbot example


I’ve created an example how to integrate pretrained model called Blenderbot to extend Rasa’s chitchat skills. It is not safe for production, but fun and unexpectedly good. Feel free to share your own ideas to make Rasa better. I hope it will be useful for some pet-projects or production chatbots with loyal customers :grinning:


hi Artem, We have been facing issues with RASA wrt training models. We decided to pre -train a model on the same data outside RASA and deploy to RASA framework in production. Is there a good way to do it ? We need the same in Prodcution actually and you mentioned that it is not safe in production


Any generative model like Blenderbot and ChatGPT can produce harmful, incorrect, biased and just toxic answer. So I would think twice before deploy it in production. Rasa team had a good point on it in their blog post. Instead of using generative models it could be more practical to use retrieval-based. Actually rasa has component for single-turn use case. If you want context dependent retriever you have to implement custom action. I don’t know any production ready tool for context dependent retriever to integrated with rasa.