Tony, the study coach chatbot


In this experiment, we wanted to build a simple Proof of Concept chatbot using Google DialogFlow to understand whether it would be useful as an alternative way to search through the SAGE resources. We were also interested in exploring how chatbots could offer more conversational interfaces; a more interactive way to solving problems for SAGE readers.


We used Google's DialogFlow to build Tony, which does not require advanced coding knowledge. We picked DialogFlow because it is a conversational platform, it allows natural language interactions and supports voice assistants. Another advantage for using DialogFlow is that it can already match similar phrases, and enabled the management of the conversational flow through 'contexts'.

First, we designed a potential conversation for the chatbot. We structured a simple flow and drafted a script of possible phrases and intents. Within the intents (high-level category of action), we added the phrases to train the bot how to respond. For example, if the user talks about giving a presentation, Tony would ask about challenges. This information would then help Tony look for an appropriate resource for that user.

We also fed the chatbot a small list of resources, where it would look for recommendations. This was the final goal of the conversation—to lead the user to an appropriate resource.

Check the outcome in the screenshots below:

Untitled design.png


We learned that bots are not that hard to create, as long as the intention or use case is very clear. DialogFlow was relatively manageable and provided built-in templates for different scenarios, such as customer support-type interactions. It also had a rapid path to production deployment.

We learned that small talk is quite important if we wanted to make the chatbot more engaging and DialogFlow had a lot of resources to support this aspect. We also understood that the final recommendation can only be as good as the data that we provide it with.

Finally, we think there are a variety of applications that could be useful for SAGE, such as alternative searches for better content recommendations that follow a less structured and more conversational approach; and of course—customer support.


James Siddle, Kieron Brown