The Death of Building Traditional Chatbots after the Advent of LLMs

Building chatbots a few years ago involved defining the flow using the concepts of intent, phrases, actions, and many more concepts. Building a simple chatbot, even for answering simple questions, used to be a rather laborious process. Everything now changed with the advent of Large Language Models (LLMs). They give us the power to communicate with the user without the restriction of the flows and give the user a meaningful response. In this article, we will examine how both work and what changed.

What is a traditional way to build a Chatbot?

The Chatbot technology is familiar and dates back many years. Until the invention of LLMs, the typical process of building chatbots relied on Natural Language Processing (NLP) concepts focusing on Natural Language Understanding (NLU). Essentially, when a user sends a message, the system uses NLP technologies to break down the sentence, identify the ‘Intent’ of the user, and propose an ‘Action’ related to the same. It relies on the flows created beforehand by the developer. Below is an example flow created for such a purpose.

Image Credits: Freshworks

When you build a chatbot using a traditional way, if the system cannot correctly identify the intent or cannot find a relevant action defined by you in the form of flows, then the chatbot is stuck and gives a standard response that it is lost and cannot answer. The chatbot is entirely relying on what you told the system to do. There is no inherent intelligence to communicate in a subtle way that it does not know.

The laborious process of building a traditional chatbot involves creating conversation flows, understanding the use cases and what potential users might ask and defining step-by-step flows with what to do when such a question arrives from the user. It is like a swim lane approach; if anything comes out of the swim lane, the chatbot is stuck.

What changed with the advent of LLMs to create Chatbots?

LLMs are one of the game changers in the Artificial Intelligence (AI) space that help us with many things that we could never do before. They are based on the Transformers AI architecture, which Google invented. LLMs leverage the power of GPUs and Transformers to give us the results we want in a fraction of a second. OpenAI’s ChatGPT, which we all see gaining popularity, is one of the examples of such LLMs. 

LLMs are trained on vast amounts of text with billions of words, such as Wikipedia and many other content sources that are available on the internet. LLMs leverage the vast knowledge and interrelationships between text using the transformers architecture below a representation of the same.

Image credit: Magic Behind LLMs

Unlike traditional approaches to defining flows beforehand, with LLMs, you do not need to define such flows. All you need to do is feed the LLMs with use case-relevant knowledge in vector databases and set the context for the LLM. When a user asks anything, the context is analysed, a relevant answer is selected, and an excellent result is generated in conjunction with the LLM model.  

No more defining of static flows; the users are free to ask anything, and this architecture with LLMs changes the entire experience for the end user. With the dawn of LLMs, traditional ways to build Chatbots seem dead, and we are in a new arena.