Managing knowledge

One of the biggest challenges for every conversational designer is managing the underlying knowledge for their chatbot project. Working for a number of years with customers from every business vertical, we've seen a whole lot of approaches to this, from text files and whiteboard drawings to Google Docs, Notion and other advanced tools. But still, managing the knowledge for a complex conversational AI project that goes beyond a simple lead generation or F.A.Q. is a complex task that requires a lot of efforts.

That's why we decided to build an advanced knowledge management system inside our bot development platform. With the help of this system, developing a chatbot with complex conversations and managing the knowledge becomes a much easier task.

What's more, the knowledge management system is fully integrated with the chatbot development process, so you can test your chatbot directly from the knowledge management interface and make sure that the knowledge is correctly understood by the chatbot.

In this article, we will show you how to use the knowledge management system to develop a chatbot with complex conversations.

Three essential chatbot technologies

There are three basic techniques involved in building chatbots (and other conversational interfaces too):

• decision trees (visual scenarios)

• natural language understanding (intent detection)

• natural language generation (question answering)

Activechat combines all these technologies in a single interface, so that you can utilize them as needed, depending on exact requirements of your project. In this article we'll concentrate on the natural language generation aspect of conversational AI, explaining exact steps that you should take to get a smart natural language assistant with minimum effort.

Types of knowledge to use

Each conversational AI project is unique, as it has to follow the needs of a specific business and its customers. In early days of chatbot development, conversation designers had to program each interaction scenario individually, creating a framework of "skills" to respond to every customers' question. In our days, large language models (LLMs) like GPT-3 or BERT make it easy to skip this tedious process and extract information automatically, generating natural language responses to almost any relevant question that customers may ask about specific business or product. All you need to do is feed this information into your project, and we'll build and train an AI model to respond to questions based on this knowledge.

Once you do this, you will be able to use these automatically generated answers both as chatbot responses and as conversational hints for your live chat agents.

Currently Activechat supports three types of business knowledge sources:

• links to websites (F.A.Q.s, blogs, marketing websites, online shops, etc)

• documentation in PDF or DOC files (SOPs, letters, business documents, etc)

• existing conversation logs (from our own live chat engine or other customer service tools like Zendesk, Intercom, Drift, or Hubspot)

How much data do you need?

A simple answer: the more, the better. Since Activechat is using only specific data that you provide, the quality of generated responses will be affected by the volume of textual information that you upload to train the model.

For example, our own question answering bot is using the whole content of this manual, as well as our website, blog, and hundreds of previous conversation logs.

As a rule of thumb, a F.A.Q. with 10 to 100 questions will be a good start. If you are an online shop, you should definitely add your storefront URL. If your business has a blog, it can also be used as a great source of information and tone of voice.