If you don’t return UserUtteranceReverted, it will keep everything, but then you need to put fallback action into the stories
I’m sorry for misunderstanding, so if you use custom fallback without UserUtteranceReverted, from the logs above it seems to me the policy cannot recall what to do after utter_fallback which could be due to the fact that there is no similar training story, or because the intent was classified as None
If it is because of None, I would recommend to try to use OOV_token in tf nlu classifier, and add nlu examples with this token as oos intent or smth
Sorry i didn’t get that
There is a possibility to set OOV_token in intent_featurizer_count_vectors, so that unseen words will be substituted with this token
but i didn’t get that how OOV will help me with my issue,
My issue is that i have a custom action which works like form action i.e when a slot is not filled , it utters for the slot value until that particular slot is filled by the user, But this implementation fails when fallback is triggered !
Now i am not getting how this will help me resolve the issue i am facing?
As I said above it depends on why it fails, I could think of two possible problems, but it might be smth else, hard to tell
Hi All, I created a Rasa init project and immediately added the configuration for fallback,so that any irrelevant texts (say I want a coffee) will go to fallback. I added the utter_default as well. But after adding the fallback, all teh utterances goes for fallback action . for ex., Hi hello etc., whereas before fallback it was going for an appropriate intent
Rasa Configuration done is as below:
- name: FallbackPolicy nlu_threshold: 1.0 core_threshold: 1.0 fallback_action_name: action_default_fallback
I change the nlu_threshold as 0.3 /0.4 etc., but all results in same result.
Kindly some one help on this.
Thanks in advance.
rasa core doesn’t see actual utterances, it sees the result of Rasa NLU classification, so it depends on intents you have for these phrases and your stories
Can you tel me what else should be done to intent classification? so that it wont fall under the fallback mechanism. Also please suggest the threshold values.
It’s impossible to say without seeing training data. Regarding threshold values it is better to experiment, default ones are a good starting point