Interactive and Training Knowledge/models


(Yoandy Mesa) #1

Hi, i’m new using rasa, apologise me if my question sounds bizarre or i if dont use the right names/words.

Is there a way to persist and use both interactive learning and “offline” trained models in a single agent? To add some context: I have some stories and trained my bot/agent using this data, then I made some online training/learning and persisted the resultant/corrected story/conversation, which i use to re-train the bot, but then it makes the same mistakes (or some times new ones) that were made during the interactive learning, when what i was expecting is it to behave like it “learned” during interactive learning session.

any hints?


(Anders) #2

I think you have to manually merge and verify the story.md files. After interactive learning I get the final action for the bot to be -export. And sometimes an intent is created called *None. Though that is probably my fault.


(Yoandy Mesa) #3

Hi Krogsager!

Thanks for your reply.

Indeed, i have manually merged “offline stories” with the resulting ones from “online training”, then re-trained using the whole set of stories.

As mentioned before, using this procedure yields the same mistakes or some times new ones (different behavior from the interactive session)


(Anders) #4

It does get hard to keep track of similar and overlapping stories and nothing prevents them from being contradicting. The recommended approach (everyone is welcome to correct me on this) is to start out with stateless stories. I.e. short stories with a single intent/action. Then use interactive learning to make the needed stories. For simple branching you can just make one story and then copy/paste & edit that story. Finally you delete the stateless stories file and use the trained one instead.