Hi Rasa Team,
I am curious what is your advice to use RASA X for prod, if I have a lot of stories already and reviewing new conversation. How can I identify If the conversation I am currently reviewing a new story, or something I already captured. If you only have a handful of stories this works but as you scale and you have multiple people using RASA X, I am curious what is the most efficient way?
Maybe this is a feature request: Would be great if RASA X would tell me based on the conversation, that here is a new story suggestion that you haven’t captured already. This should be relatively easy as you already have all the information in the database.
Another feature request: Would be great to create a story inbox as well same as NLU inbox. Where I would see new stories that I haven’t captured yet, and I could make a decision if I want to add that story or not. Challenge here is how you define the end of starting point of the story. Timestamp would be interesting here and I as a user would be able to set for RASA X how to cut the conversations.
Basically I see 2 problems RASA X will help me to solve:
- Identify new stories
- Add more training data to NLU
Both these problems have 2 decision points for a RASA X user: a. Being able to know exactly the current state of NLU / stories b. Based on this knowledge being able to identify what is a new data, that you can add to NLU / stories
RASA X can solve the problem, if it can solve a) and b) problems the easiest way.
NLU currently does a good job with the inbox, because it shows me new NLU content and I can decide if I want to add it or not (however what I have experienced is if I review something in NLU inbox, next day it will appear again. I don’t know if it is a bug or by design)
I think if there would be a Stories inbox, which helps me the above mentioned a) and b) problems, I would be able to identify new Stories easily. However solving problem b) for stories is not trivial, as a definition of a story is not as straightforward as a definition for an intent / NLU element.
Story has two properties: starting point and end point. This is the tricky part, because RASA X at processing time won’t know from a conversation how to identify a story. If my bot has a 100 turn conversation with a user, than the different story permutations are huge.
So that’s why I am curious how people review stories and can keep in mind what stories are already exist in the yaml files and what are new stories that we would need to give the model to train.
I’ll think more how would be an efficient logic to identify stories and share if I have any ideas.
Keep up the good work!