I’ve read the docs, and I’m wondering if anyone from the Rasa product team can tell me if there are any plans to add specifics for handing off to a human to Rasa Open Source.
It seems to me that the ability to hand off is vital for operating bots in production. Research on the negativity bias suggests a 4:1 good to bad ratio is required to keep people coming back. That means for every one bad experience with a bot, the user must have four good experiences otherwise we’re going backwards. Bots are still new, and most people are willing to give them a try. That willingness could erode if we’re not careful.
The word “artificial” in AI is a cliché. Amplified or augmented is much more appropriate. The goal shouldn’t be to replace humans but rather to augment them. Think orthoses, like reading glasses.
The question then becomes, how best to bring the human back into the loop? Two suggestions:
- Trigger an event on low NLU confidence; a
ThreeStageFallbackPolicythat enables a human third party to join an existing channel and see the entire conversation history. - A slight modification of the interactive learning mode where a human third party confirms each bot response or intervenes.
Does anyone here have experience running Rasa in production with real users and a human in the loop? How did you do it?