In this repo, I will show how to build a simple chatbot for the change plan use case using Rasa NLU and Rasa Core libraries. In particular, I will show how to build such a system based on a workflow that we designed before. The ChangePlan flow is shown below:
The ontology design and data collection are largely based on the flow.
The ontology is defined in the changeplan_domain file. There are five user intents and six bot actions. Two slots are involved: planType and isGrandfathered. We also define some prompts.
The training data for NLU and policy learning are available at data/nlu/data.json and data/stories.md respectively. The NLU annotations are manually generated while the policy annotations are created by rolling out the workflow.
You can use the repo to build a chatbot within a few minutes! Two example conversations are shown below. They differ in the is_grandfathered slot value.
Conversation 1 (is_grandfathered=True):
Conversation 2 (is_grandfathered=False):
Install Rasa libraries with:
pip install rasa=0.15.0a6
If this does not work, then:
pip install rasa-core==0.13.2
pip install rasa-nlu==0.14.6
To train and test the NLU model run:
python nlu_model.py
(you may train multiple times until "yes" leads to the confirm intent...)
To run the dialogue manager, we need to start a server that is configured in a 'endpoints.yml' file. This is how to train and run the dialogue management model:
- Start the custom action server by running:
python -m rasa_core_sdk.endpoint --actions actions
- Open a new terminal and train the Rasa Core model by running:
python dialogue_management_model.py
- Talk to the chatbot once it's loaded.
We can use the Rasa interactive training feature to collect more training data for both NLU and policy learning. The process of running the interactive session is very similar to training the Rasa Core model:
- Make sure the custom actions server is running:
python -m rasa_core_sdk.endpoint --actions actions
- Start the interactive training session by running:
python train_interactive.py