I am using Rasa 2.8 (it is old project) And i have 258 intents, 322 Entities, 50000 examples. and 589 stories. Customer delete some stories and add some few other stories. Then trained the model. And now whenever they ask question related with intent which doesnt have action it answers perfectly. And after few this type of questions they ask another question which related with intent which has action it does the action perfectly. But when ask a question with intent which doesnt have action it fall into fallback case of policy. It start to say “Do you mean this” and give me an intent name. Here is my policy : policies:
-
name: TEDPolicy max_history: 10 epochs: 300 constrain_similarities: true batch_size:
- 32
- 64 embedding_dimension: 256 rnn_size: 256 dropout: 0.2 weight_sparsity: 0.8 evaluate_every_number_of_epochs: 20 gradient_clip_value: 5
-
name: rasa.sf.core.memoization.MemoizationPolicy
-
name: rasa.core.policies.form_policy.FormPolicy
-
name: rasa.core.policies.rule_policy.RulePolicy core_fallback_action_name: action_default_fallback core_fallback_threshold: 0.3
when ask grok or chatgpt it offers me to add some stories which include not just actual dialog for an intent but also include continuation actions like chitchat intents etc in a dialogue. And also offer me the increase core_fallback_threshold to 0.6 . I made them and trained the model once more. But this time situation get worse.
by the way at my local server when i trained with initial pipelines and data i couldnt catch their case. At local when i tried questions without action model was answering and whern i use questions with actions it did the action and return me the custom utter which is prepared by action code. and also give me fallback utter too.
When i made chatgpt offers and train again now it almost always fall into fallback case in answering part. Anyone can offer me something which can help me ?