Some History of Support Bots

Despite seeming like a futuristic idea, chatbots have been around for over fifty years. One of the earliest chatbots, a psychotherapist emulator nicknamed ELIZA, was created back in 1966 at MIT. ELIZA could interpret text and respond to prompts based on a script. And though ELIZA captured the interest of contemporary programmers, academics, and futurists, it wasn’t until a bot called SmarterChild (created by ActiveBuddy, Inc.) was integrated into America Online Instant Messenger (AIM) in 2001 (and later integrated into MSN Messenger) that the public could interact with this type of technology. You could ask SmarterChild for the weather, sports scores, movie show times and more. At its peak, SmarterChild had over 30 million AIM buddies and accounted for almost 5% of the entire Internet’s chat volume.

Leaping forward to 2023, chatbots have become more sophisticated as advancements in technology have allowed them to efficiently and realistically mimic human interaction through complex algorithms and machine learning. Although they haven’t quite passed the Turing Test yet, they are certainly getting closer.

NextLevel Support Bot Effort

As a self-funded startup company, Nextlevel Mediation was looking for ways to streamline operations and still offer superior service and ensure user satisfaction. One remarkable tool that is having an impact in digital space is the Technical Support Bot. Having integrated AI into various aspects of our mediation software platform, it made sense to expand its functionality into the mediator technical support space. The basic requirement for the support bot was to help mediators navigate the platform and resolve questions about key functionalities with no human intervention, as well as be available 7/24. In addition, it had to be trained to provide general help with mediation questions.

A well-structured technical support bot deflects not only the load off our customer support staff but also enhances our user’s experience. Also, it frees our business from the constraints of time zones and working hours, providing constant, instantaneous support.

So, what were the key requirements that went into building the NextLevel technical support bot?.

• User-Friendliness: If a bot is hard to comprehend or interact with, it defeats its purpose. So one of the key requirements was to make sure that the bot had a user-friendly, intuitive interface that allows users to easily seek answers to their questions.
• Problem-Solving Capabilities: To effectively address user queries, the support bot had to have built-in problem-solving capabilities. It had to understand what the user wanted and provide appropriate solutions based on its semantic long term and short-term memory training.
• Knowledge Base Integration: The long-term memory training had to be integrated into a comprehensive knowledge base and dynamically draw on this information to address queries and guide users through important functions.
• NLP and Machine Learning: Natural Language Processing (NLP) and Machine Learning (ML) were essential for our bot so that it could understand and react to human language in a dynamic context. In addition, the bot had to continually learn from previous interactions (at least within a session) as well as handle multiple users.
• Self-Service Accessibility: While the bot did not have to create support tickets when problems were encountered, it had to guide users to the NextLevel self- service portal where they can create their own tickets if their issue remains unresolved.
• Scalability: The bot had to be scalable, so that as the business grows and the number of support questions increases, the bot can handle the activity load.

Challenges is creating a Technical Support Bot

• One of the major challenges using Generative AI as instructional support bots is non-determinism. Non-determinism, in LLMs, means that the model can produce different outputs even when given the same input. This behavior is a byproduct of the complex neural networks and vast amounts of data used to train these models. While non-determinism can lead to creative and diverse outputs, it can also cause inconsistency. Inconsistency can lead to confusion and frustration for our users. Imagine asking the support bot how to do something in the system and getting a different set of instructions each time.
• Another challenge is that LLMs do not do logical reasoning like humans. According to a research paper by scientists at the University of California, Los Angeles, transformers, the deep learning architectures used in LLMs, don’t learn to emulate reasoning functions. Instead, they find clever ways to learn statistical features that inherently exist in the reasoning problems. That means the support bot has to be implemented with an additional set of skills for complex support questions.
• The last challenge is generating an effective prompt. The prompt must contain the right balance of intent, relevant semantic memories, historical conversations, and token management in order to achieve an optimal response. Extracting user intent in a way that AI systems can use along with the structuring of vector memories for accurate retrieval, are key factors that play into this challenge.

While these issues are related to the quality of the response, there are other issues related to performance. Having a relevant instructional response that takes too long is not good either. This can also be a major cause of frustration, like being in a long waiting queue for telephone support response. However, this optimization is more closely related to the choice of vector databases, temperature and p-stop parameters, and the number of semantic searches performed when answering a question.

In conclusion, creating a technical support bot that helps the user to use software and answers questions about key functions is not just a luxury, but a necessity in today’s business environment. A well-crafted bot can become one of your greatest assets, promoting user satisfaction and increasing overall efficiency. However, overcoming some of the design challenges takes a lot of experimentation.