This week, we made significant progress on our React configuration web app.
Configuration web app and Redux
We integrated Redux into our web app to allow users to generate a config file for AskBob with their customisations. The web app allows users to specify intents (questions or commands the user says), responses (Bob’s replies to the user) and make skills (an intent paired with a sequence of responses), and then download a config file containing the data. This file will then be able to be used to configure the AskBob voice assistant.
When the user adds an intent, response or skill in the web app, the data is stored locally to allow users to leave and return to their configuration a later date. Redux is used for state management as the state has become quite complex (and may grow in complexity once forms are added) and Redux gives us the flexibility to scale our state even further.
The lists of intents, responses, and skills are displayed on their respective pages, allowing users to quickly see their current configuration elements. Users can also edit and delete intents, responses, and skills as well. These changes are instantly stored in local storage.
The forms used to add data to the configuration file have validation checks to make sure users do not add badly formatted data to the configuration. This is crucial as all data to be used as training data for the Rasa NLU model must be correctly formatted. All responses and skills must have at least one variant (way of saying it), so the form will notify the user if they do not specify one. Also, the form will notify the user if they have a response ID that doesn’t start with utter_ (as required by Rasa) or any empty fields. As a result, the user cannot add badly formatted data to the config and these errors must be fixed before a configuration file can be generated.
Converting the configuration file into Rasa YAML configuration files
On the voice assistant side, we wrote a Python script to convert a config.json file exported by the configuration web app into the series of files required by Rasa for training. This allows us to take exported configuration files, run the conversion script and then run rasa train in the command line to train a customised NLU model. This can then be launched by a Rasa server and immediately start processing transcribed spoken requests from AskBob.
Next steps
One of our next steps is to change our forms to use a more user-friendly drag-and-drop interface. This will allow users to create intents, responses, and skills on the same page in a more intuitive manner. We aim to implement a Scratch-like block-based customisation interface to allow users to make more complex voice skills for AskBob.
We will also make a unified setup script for AskBob to automatically train a Rasa NLU model after converting a config.json file and lay the foundations for an AskBob plugin system.