Is there any example on fallback actions ??
Hey @aby09. We have an example of implementation on the docs Fallback Actions. In addition to that, you can check out Sara, the demo bot which uses fallback policy as well: GitHub - RasaHQ/rasa-demo: Sara - the Rasa Demo Bot: An example of a contextual AI assistant built with the open source Rasa Stack
okay i ll check out SARA … thank you
how to set threshold value for nlu and core to get right answers ?
trial and error method or any other methods?
you can mention threshold in your core or nlu training scripts like this
fallback = FallbackPolicy(fallback_action_name=“action_fallback”, core_threshold=0.2, nlu_threshold=0.3)
agent = Agent(‘domain_dialogue.yml’, policies=[MemoizationPolicy(), KerasPolicy(), fallback])
and pass the fallback in your agent with other policies like above
how to set value ?
if you are talking about confidence value then its like that
how you came know that 0.2 and 0.3 are best values for ur chatbot ? i want that …
thats not for your chat bot… that config can be changed according your need… i added that for my own config… use some sense pls!
dude i know that … after setting up threshold value i am not getting proper answers. So i am asking is there any method to know proper threshold value
i guess you are setting nlu threshold too high which pipeline you are using spacy or tensorflow??
i am using spacy
Even i faced the same problem, it was quite tiresome to frequently change the threshold values and again train the model so is there any relevant method to find a proper threshold value.
even i am facing same problem …i guess you should do it manually
someone plz prove me wrong …
No you don’t have to change it everytime… See it depends on your training data… The nlu threshold you have entered is too high 0.8 is very high… do one think ill prefer changing your pipline to tensorflow… y tensorflow? cz tensorflow works good for intent classification with higher confidence with low training data as compared to spacy or mitie… you have to have the knowledge for correct pipelines in this domain and for that you can refer in the docs…They have a good collection for docs you just have to sit and go through for better understanding… I have right now many bots working on rasa NLU and Core and i didnt have to change the threshold for any of them…just train them properly and go through the pipeline functioning…
And by lower training data it doesn’t mean 5-10 phrases per intent… You still need at least 15-20 or max 25 for tensorflow and for spacy you need far more than that for better confidence
its unethical comment
@aby09 This blog might be of help to you in finding the apt NLU threshold. https://blog.rasa.com/failing-gracefully-with-rasa-core/?_ga=2.240287815.518918257.1580717424-419595417.1574159216
for version 2 it is not working using FallbackClassifier
i have added more details on