HI! (first sorry for my bad english…:)) and probably this is most suitable section to post; sorry for my previus post in in the wrong forum section;
I’m Rasa’ newbie, and I started studying it again a few weeks ago :). But I’m not new to NLP and DeepL (I’m work almost daily with TF,Pytorch, Python and R…). but I have a question for RASA senior and developers:" can be suitable (ther’re already some API about…) to integrate chatGPT in Rasa workflow?? I think that today “unfortunately or fortunately” in all NLP (mostly in chatbot systems)it cannot be ignored, or better: cannot be ignore it’s architecture:
transformer decoder trained by using human feedback to guide the learning process with the specific goal of mitigating the model’s misalignment issue: → Reinforcement Learning from Human Feedback.
For example, chatGPT actually is able to extract all “intents” in many languages without “fine tuning” (undoubtedly this is a very powerful feature), and ubsequently it will be possible mapping they with intents of our interest, with a simple similarity algorithms; it actually is also able to extract all kind of NER from text (without fine tuning)…and so on… why not take advantage of this new “architecture” in – RASA + RLHF --?? (in this way maybe , there would no longer be a need to manually write all possible intent’s exampels, synonyms, entities, training NER extractor, and so on…)…could it be so???
i meant to include a chatgpt like model in Rasa core…an almost total upgrade of Rasa 's architecture would be needed, because with new LLMs algorithms probably Rosa would no longer have reason to be conceived in that way…
in a while chatGPT’s like algorithms can be fine tuned on ANY TASK…(with appropriate external associative memory), then will be no need any intent or NER retrieval…
then actual RASA architecture will be useless and obsolete very soon…this is the reality of the facts
You may want to check out my GRACE project that explores just this idea. It doesn’t use Rasa but instead relies on GPT-3 to implement a chatbot capable of executing code/commands and fetching information from external sources.
then actual RASA architecture will be useless and obsolete very soon…this is the reality of the facts
Despite how much Rasa has advanced over the past years, I tend to agree. The future is few-shot, not fine-tuning.
I believe that the pro version already has some kind of functionality with GPT, but I found an interesting project that I’m testing. https://botpress.com/
You can try reading this article: How to use ChatGPT with Rasa - Python Warriors
It helps to setup ChatGPT support in case of NLU_fallback from rasa model. You can also tweak some changes to make it available for other cases as well.
