Advanced Knowledge Understanding Chatbot, Introducing DeepKU - Pt 1

TecAce Software
Aug 21, 2023
4 min read

We live in an era overflowing with new information and knowledge. Among them, the ability to find and utilize information that we’re curious about or need is increasingly establishing itself as a crucial competitive advantage.

Until recently, we relied on web searches, but now, with the emergence of chatbots based on large language models like ChatGPT (LLM-chatbots), the way we seek information is changing. This is because you can now directly ask LLM chatbots in natural language if you have a question.

Advantages and Disadvantages of Web Search

Until now, if someone had a query, they would define search terms that might contain the answer and then search on platforms like Google, Bing, and Naver. They'd read the resulting links to find their desired answer. Given the vast amount of diverse information on the web, the likelihood of finding what one is curious about is high.

However, web search has disadvantages, as shown in Table 1. The first is that to quickly find the desired information, one must craft effective search terms. Second, because there's a plethora of information on the web, choosing the most suitable document from numerous search results isn't straightforward. Lastly, web searches provide documents with high relevance but don't necessarily present the desired answer. To obtain the desired answer, one has to invest time and effort in reading and deciphering the provided documents.

Table 1. Disadvantages of Searching

Disadvantage	Description
Keyword Generation	To find the desired information, it's crucial to craft the right search term.
Selecting Search Results	There are numerous similar matches in the search results, and the right one must be identified among them.
Crafting an Answer	After reviewing the search results, one must derive the answer they seek.

Advantages and Disadvantages of LLM-Chatbot

With LLM-Chatbot, one can directly ask in natural language what they're curious about. Since it provides answers after reading multiple documents, it's more convenient than searching.

On the other hand, it's important to consider that the LLM-Chatbot might have difficulty or even be unable to provide answers to the latest information or content it hasn't been trained on.

Additionally, since LLM-Chatbot is based on neural network-based learning, the abstracted information may differ in detail during the conversion to an answer. This means it may sometimes provide incorrect or strange answers due to abstraction errors.

Table 2. Disadvantages of LLM-Chatbot

Disadvantage	Description
Dependency on Trained Data	It's hard to answer about the latest information or content that hasn't been included in its training.
Characteristics of Abstracted Information	Due to the abstraction process in the neural network training, the response might be incorrect or seem bizarre.

The emergence of Web-based Chatbots

To overcome the disadvantages of the LLM-Chatbot, one can use web search-based chatbots (WebChatbots) like BingChat. This method extracts keywords from the user's question, searches the web for relevant content, and then converts the content to a response using LLM. This allows the chatbot to provide answers even to the latest information not covered by LLM.

WebChatbots combine web search capabilities with LLM-Chatbot, making them useful in various search situations. However, since the chatbot hasn't been trained on the web-searched information, there's a possibility it might not fully understand and might generate an answer based on abstract knowledge. For general information queries, it might be better to use the LLM Chatbot.

Advanced Knowledge Understanding Chatbot, DeepKU

The advanced knowledge understanding chatbot that AceAI is pursuing, DeepKU, aims to overcome these disadvantages. It aims to provide accurate answers based on a deep understanding and utilizes both the latest external data and its inherent information.

In this article, we'll introduce one method of addressing the mentioned issues by using the option features. In subsequent articles, we'll delve deeper into fundamental solutions and showcase how the DeepKU methodology is applied in real-world scenarios.

Currently, DeepKU is offering services based on the GPT LLM and the DuckDuckGo search engine. However, it's possible to change to any desired service based on LLM or a different search engine.

Upon accessing DeepKU's LLM-Chatbot service, users will encounter an interface that looks like Picture 1. Notably, DeepKU's Strategic Chatbot employs a unique AI algorithm, different from conventional chatbots, which operates in a way that contributes to achieving the strategic goals and objectives of the organization it serves.

Picture 1. DeepKU Login

DeepKU's strategic chatbot (see Picture 2.) is designed to flexibly adjust three core parameters either at the beginning or in the middle of a conversation. This minimizes the limitations of information retrieval based on the drawbacks of LLM-Chatbot and WebChatbot.

Picture 2. Adjustable parameters in DeepKU’s Strategic Chatbot

It includes a slider selector that allows users to adjust the "chat temperature" of the chatbot. If users want factual answers, they can set it closer to 0. If they wish for detailed, albeit potentially exaggerated explanations, they can set it closer to 1.
Users can change the AI model used for chatting. Although each interaction feels continuous to the user, the system actually sends all the previous data to the AI model to retrieve the next response. Even if the model is changed mid-conversation, the flow isn't interrupted. This allows for better efficiency, especially when longer conversations are needed.
The most crucial real-time adjustable parameter is the inclusion of web search. Users can activate web searches at any point if they wish to pull answers from a broader knowledge pool or the latest information. With this feature, they are more likely to receive concise, factual, and simplified answers. If they want detailed responses based on comprehensive knowledge, they can deactivate the web search feature.

Conclusion

In this article, we looked at the increasingly utilized LLM-Chatbot and WebChatbot for information navigation. We also explored the flexible chatbot, DeepKU, which can seamlessly combine both. In subsequent articles, we'll discuss the advanced features of DeepKU and its innovative algorithms and applications.

🇰🇷 Here is the Korean version of the article: