This week, we made additional progress with improving the customisability of AskBob.
Rasa model training from AskBob build-time configuration JSON files
The Python script mentioned in last week’s blog post to translate AskBob “skills definition” build-time JSON configuration files into the necessary YAML files required to train and configure Rasa has been developed into a setup utility accessible on the command line through python -m askbob --setup config.json. It now allows users to train and output a new Rasa model that is immediately available for AskBob to use when it is next launched.
We also continued experimenting with the format of the JSON configuration files to ensure it is flexible enough for us to be able to add additional features in the future beyond those we have already implemented while resolving compatibility issues between the voice assistant setup utility and our configuration generator web app.
AskBob query responder exposure via a RESTful web API & FISE ecosystem integration
AskBob’s query response logic has been exposed through a RESTful web API to enable other teams and developers to use AskBob as a service. If AskBob is run with the -s or --serve flags on the command line, then a Sanic-based Python web server replaces the speech-based interface. The host and port to which the new “AskBob server” binds is configurable within the runtime config.ini file.
Besides a welcome on the root / path, the API has a single HTTP POST /query endpoint taking a message for the assistant to parse and respond to, as well as a sender (a unique identifier used to keep track of the conversation) sent using the application/x-www-form-urlencoded content type and returning a JSON response.
If the request is successful, then the server will return the response as a sequence of messages:
{
"messages": [
"Hello, world!",
"I'm Bob, how do you do?"
]
}
If an invalid request is given, then the server will return an error taking the following form:
{
"error": "Error message."
}
We also held another round of meetings with the other two IBM FISE teams this week to discuss our strategy for integrating our projects. The video conferencing group should be able now to host AskBob on their server and then access AskBob through their back end using the /query endpoint. They are already using IBM Watson for speech-to-text and text-to-speech, so should be able to send the transcribed speech directly to AskBob and use the JSON response to produce audible feedback for the end user.
Configuration generator webapp
We have also investigated a few libraries that would help us build drag-and-drop components for the AskBob Config app. We looked into Blockly, which would allow us to build a Scratch-like drag and drop editor, using blocks that could attach to each other. However, we found Blockly not necessarily suitable for what we needed. Blockly allowed for code generation and provided a customisable user interface with customisable coding blocks, but it was hard to integrate it into the React app. Also, a lot of the features it came with were not necessary for what we were building.
Sortable was another library we looked into that turned out to be more suitable. Sortable allowed us to quickly make a drag-and-drop list of intents and responses and easily change the order of items in that list. Sortable was used for both the skills and stories form, improving the UI in both sections.
Next steps
Now that we have a unified setup utility for AskBob, we will look into developing a plugin system that will enable us to modularise assistant skills into separate packages loadable at build-time (the time python -m askbob --setup config.json is executed to retrain the Rasa model).