On Chatbot Prototyping

Last year we got into chatbots and conversational AI services. We learned about many relevant solutions in the industry and got to know some of the companies behind them. Recently got to propose a chatbot related project to a potential client, including some prototyping as part of the pitch. While the client is considering the option, and comparing our pitch to competitors, we felt like blogging about the experience and go into what we proposed and what made the prototype special.

A chatbot interviewing a human by midjourney

A key function that this chatbot needed to have was to interact with an existing service for exchanging data. This service is run by a separate company and the means of communication is via proprietary XML formats. I.e. the chatbot needs to be able to read their XML messages, dynamically create conversations from them, and return what it learned through the conversation back in a new XML message. As per clients request we opted to use the OSS botpress for our planning and prototype. The proposed work packages included roughly,

• General chatbot behaviour (greeting, intent recognition, general knowledge, ...)
• Interaction with XML service, ie. receiving XML, dynamically parsing them into conversations, creating XML responses from user interactions and sending them to API.
• Consulting, i.e. the chatbot needs to consult users into what they want to do in the first place. This requires a RAG solution in the back.
• Deployment & testing, we usually prefer to offer solutions in cloud, but we also show that on-prem is a possability (e.g. using ollama). 

System sketch of proposed prototype.

The basic XML interaction we implemented as the prototype to go along with the pitch. We showed how the XML is turned into a conversation, how that conversation is enhanced by running both the chatbot texts as well as user responses through a LLM, and how the provided information is turned back into an XML. We got to show the conversation flows in botpress studio as well as its configuration capability. On our side, we had implemted the prototype using python scripts for reading the .xml files and creating the .flow.json files, which botpress is using under the hood. Both implementing and presenting the prototype was fun, and we are excited to see how the client will decide.