How to Make ChatGPT Smarter: Part 1
Nowadays, everyone has heard of ChatGPT, which is an LLM or Large Language Model that seems to generate text very much like a human would. However, if we look deeper, the AI of Chat GPT works by probabilistically selecting the most suitable words to respond to us. If we want it to respond exactly as we desire, we can use a Prompt Canvas to help in asking questions. However, in this article, we will look at ways to make Chat GPT even smarter through two methods: RAG and Agentic AI.
Method 1: RAG
RAG, or Retrieval Augmented Generation, involves inputting your own data into the Chat GPT model and querying information from the files or data you have provided. RAG and generative AI – Azure AI Search | Microsoft Learn
For example, you can input financial statements through Azure Open AI and query information from those files. The advantage of this method is that you can use your own uploaded data. However, the downside is that setting it up in Azure Open AI can be quite complex, and it only supports certain types of files (for instance, as of October 9, 2024, Azure Open AI does not yet support Excel files and information in SharePoint).
Method 2: Agentic AI
Another way to make any AI smarter is through the implementation of Agentic AI. The concept of the Agentic Model was introduced by Andrew Ng to enhance AI performance by using multiple-layered AI models instead of a single layer. For example, using GPT-3.5 in conjunction with the Agentic Model yields better results than using GPT-4.0 alone. This method allows the AI to correct its own errors. Additionally, system messages in the GPT model can be used to define the desired characteristics of the AI. For instance, one AI could write code while another checks it, or one AI could plan tasks before executing them step by step. This approach enables the AI to work more efficiently and flexibly. (Ref: https://www.facebook.com/share/p/zpPjbN4PxPepYzna/)
In this article, we will explore ways to make Chat GPT smarter through the following four experiments:
- Single-round Questioning with Chat GPT, also known as Zero-shot Prompting. The advantage of this method is its simplicity. However, as Andrew Ng has pointed out, when we type, we often use the backspace key to make corrections or review our work before sending it. Similarly, AI should also review and correct what it generates.
- Ask Chat GPT and 2-3 Others, Then Summarize and Choose the Best Answer. This method involves using Zero-shot Prompting, similar to the first method (i.e., asking in a single round). However, instead of asking only Chat GPT, you can ask multiple models, such as Chat GPT, LLama, and Claude. You then combine and summarize the answers to determine which one is the best. Of course, this approach yields better answers than the first method.
- Ask Only Chat GPT, but Have It Review Its Own Answers. In this method, you ask Chat GPT to review its own answers and consider whether a user would be satisfied with the response. If not, it should revise the answer accordingly. You can use a System Message to facilitate this process. For example, you might write: “You are the logic and reasoning engine who will revise the answer and make sure the answer is coherent and makes logical sense. From experiments, this approach, even though it uses only a single model, yields better results than the second method. This is because the second method does not involve self-correction, whereas this method entails multiple rounds of self-review to improve the response.
- Ask Multiple Models and Multiple Rounds. This method combines the second and third approaches. You start by asking Chat GPT to reflect on its answer repeatedly until the response is satisfactory. Then, you move on to Claude, having it reflect on its answers in the same manner, and finally do the same with LLama. Afterward, you ask Chat GPT to summarize the final answers from all three models into one final, coherent response. Although asking multiple times across several models provides the best performance, it comes with high costs. These costs include the expense of multiple API calls (3 models * 4 rounds) and consuming more nodes/steps in your quota on platforms like make.com. Additionally, the complexity and number of steps increase the risk of failure due to the intricate process.
Of course, this approach yields better answers than the first method.
You can use a System Message to facilitate this process. For example, you might write: “You are the logic and reasoning engine who will revise the answer and make sure the answer is coherent and makes logical sense. 
From experiments, this approach, even though it uses only a single model, yields better results than the second method. This is because the second method does not involve self-correction, whereas this method entails multiple rounds of self-review to improve the response.
Although asking multiple times across several models provides the best performance, it comes with high costs. These costs include the expense of multiple API calls (3 models * 4 rounds) and consuming more nodes/steps in your quota on platforms like make.com. Additionally, the complexity and number of steps increase the risk of failure due to the intricate process.How to Make Chat GPT Smarter with Function Call: Part 2
Summary
We recommend using the third approach if you opt for Agentic AI. This involves selecting a single model and having it review its own answers about 2-3 times before finalizing the response. Additionally, when implementing AI, you don’t need to stick to just one method. You can combine different approaches, such as integrating RAG with Agentic AI and others. These strategies for enhancing AI intelligence and selecting the appropriate solution are the tasks of Data Alchemists.
If you need a consultation with an AI Alchemist, feel free to contact us via Line: @davoy
Subscribe to get great articles about data
[hubspot type=”form” portal=”47406981″ id=”244b28eb-5322-4f57-80ff-1745ac75313c”]