Rasa fallback policy

I am working on a project. I have created a chatbot with rasa open source. However, I not able to use fallback policy properly.

I have used out_of_scope but it doesn’t serve well. I would be grateful if the chatbot can return the unknown answer with a specific answer like "Sorry, couldn’t understand that! can you rephrase.

There is not complete tutorial on that. I am wondering why nlu.yml is not touched there, Fallback and Human Handoff

Hi, can you show your current implementation?

Can we have a video call on google meet to show the work? my email tashinus@gmail.com

Hi @tsheringtashi, you can post your code or link to GitHub here so others will be able to help as well.

Thanks, I have link embedded in Linkedin. I would like to know how to operate handoff and fallback policy.

  1. Uncomment the pipeline components in the config file

  2. There are many stories that can be written as rules (to follow strict behavior)

I have uncommented the pipeline components in the config file. However, the problem remains.

I wanted the bot to answer even if there is not intent in nlu.yml.

below for instance:

User: “who is the president of China?”

Bot : “Sorry, I couldn’t understand. can you rephrase it?”

You can add a rule for the Fallback. Here you can find how to implement it.

I am sorry for not understanding Rasa properly.

Do I have to use only rules from rules.


  • rule: Ask the user to rephrase whenever they send a message with low NLU confidence


    • intent: nlu_fallback

    • action: utter_please_rephrase

Basically, Stories include multiple conversation paths and patterns (like receiving various info from users) and Rules follow strict behavior (like greeting or asking the user to rephrase)

Thank you, it is working but not meeting the expectation.

Again explaining with the used case below:

User: Who is the president of America?

Bot: Bye

What should I do?

Here the bot is predicting the user message with the wrong intent. You can use rasa shell nlu to check the confidence of the intent prediction. Based on that you can play with the threshold in config file to meet your needs.

    - name: FallbackClassifier 
      threshold: 0.7 # play with this

thank you. How to deploy the application in docker?

I have uncommented the pipeline components in the config file. However, the problem remains..


Hi @khaanekii, please create a new post regarding your problem