FAQ Bot Policy

I’m working on making an FAQ bot and I was wondering what kind of policy I should/need to use. Every story is incredibly straightforward, e.g.:

## question_1_story
* question_1_intent
  - utter_question_1_answer
  - request_feedback

## question_2_story
8 question_2_intent
  - utter_question_2_answer
  - request_feedback

Since this is all so simple, do I even need something like the Keras policy that uses ML? Do I need to create my own custom policy? ML just doesn’t seem necessary and could make the bot worse since it’s probabilistic, not deterministic.

Hey @mjspeck. You’re right, if it’s super simple, deterministic is better. However, if you have multiple policies in your configuration, there is an implicit prioriization that goes on behind the scenes, so if there is a memorized action, the machine learning policy won’t be invoked.

What I would suggest is testing your bot without ML and making sure that it answers all the paths you want correctly. Then if you add KerasPolicy, it most likely won’t be used, but if some weird context shows up that your bot hasn’t seen somehow, it can hopefully help steer toward the right action.

1 Like