Bot don't respond

Hello, Need help please.

My bot respond with correct answer during the online training, and when I confirm the answer he want to give, it works (use a custom action to asnwer). How ever when I try my bot with actual dialogue It don’t work.

the sénarion will be like: User: Bonjour Bot: Bonjour écrivez votre problème User: mon probleme est: bla bla Nothing from bot

In my stories file I have something like:

## Generated Story -960033508281559891
  • bonjour
    • utter_bonjour
  • confirmer
- utter_etape1
  • problème_site
- SendIntent_probleme_site
- utter_problème_site

So according to my story file the bot should do the action SendIntent_probleme_site. The class for this action is: class SendIntent_probleme_site(Action): def name(self): return ‘SendIntent_probleme_site’ def run(self, dispatcher, tracker, domain): message = tracker.latest_message.text #print(message) return [SlotSet(‘intent_probleme_site’, message)]

I do this to store the intent_probleme as a slot to use it later

Any help please? I also don’t know how to do something that tell me were is the problem like a log or exception, etc.

Thank you

You can pass the --debug flag on running rasa core to get debug of what is going on in each step of the conversation. it will be useful to notice the logs for this problem

Thank you responding here is what I get in the log gile this is when the bot do not respond (nothing)

[11/26/2018 10:21:36] DEBUG:There is no memorised next action
[11/26/2018 10:21:36] DEBUG:Predicted next action using policy_0_KerasPolicy
[11/26/2018 10:21:36] DEBUG:Predicted next action 'action_listen' with prob 0.55.
[11/26/2018 10:21:36] DEBUG:Action 'action_listen' ended with events '[]'
[11/26/2018 10:21:36] DEBUG:Current topic: None

Your example don’t match your story

You said User:Bonjour(intent:bonjour) Bot: Bonjour, ecrivez votre probleme(action:utter_bonjour) User: Mon probleme est xyz(intent: confirmer ou probleme site??)

Your story is trained on something else than you example

Actualy I just didn’t put all the sénario I Used to test. But for testing I had the same sénario as in story. And as I said in the online training the same testing sénario worked fine. The bot find the right response. But in the final testing the same sénario do not work.

The real testing sénarion:

User: Bonjour (intent: bonjour)

Bot: Bonjour, vous avez émis une idée de progrés. Voulez vous que je vous aide a emettre d'autres idées? 

User: Oui (intent: confirmer) 

Bot: Ecrivez a nouveau votre probleme (utter_étape1)

User: plusieurs personnes ne participent pas (probleme_site)

Bot:                         (bot do nothing instead of doing SendIntent_probleme_site and then 

What are your training parameters? which policy did you use for training the model

It seems that the bot is not able to follow the story based on your logs, as it says there are no memorised action and calls the Keras policy instead which gives a wrong prediction. Keras policy will need a lot more training data and also to see how your parameters look like for training the models

Did you retrain after online training? if finetune is not true, you should retrain the bot

This is my config file:

language: "fr"

- name: "intent_featurizer_count_vectors"
- name: "intent_classifier_tensorflow_embedding"
intent_tokenization_flag: true
intent_split_symbol: "+"

And yes I do retrain the model after online training. I tried to upgrade to test, and now I have to change a lot of things in my code that are not supported any more. Will try like this and see

I meant the config of Rasa core. Your rasa nlu seems to be working

Here, I just took an existing example to use it in my case:

agent = Agent('chat_domain.yml', policies = [MemoizationPolicy(), KerasPolicy()])


Which version are you at? 0.11 ?

Since you are using only tensorflow, you don’t have any entities either that sometimes causes issues.

It is strange that the defaults are not working.

How many stories you have got? maybe make augmentation to 0 and reduce the epoch to 100 ( not sure if your NN will anything after 200, if you have reached your minimum loss)

It is hard to asses like that without knowing the full picture. You can try to evaluate the rasa core model as well