2023-11-13-How to Build a Personal Knowledge Base Application in 5 Minutes Using Natural Language? My GPTs Builder Attempt - Minority

How to Build a Personal Knowledge Base Application in 5 Minutes Using Natural Language? My GPTs Builder Attempt - Minority#

#Omnivore

How to Build a Personal Knowledge Base Application in 5 Minutes Using Natural Language? My GPTs Builder Attempt

Matrix Homepage Recommendation

Matrix is a writing community of Minority, where we advocate sharing genuine product experiences and practical insights. We periodically select the highest quality articles from Matrix to showcase the most authentic experiences and viewpoints from users.

The article represents the author's personal opinion, and Minority only makes slight modifications to the title and layout.

Once the developer's imagination gate is opened, what awaits us is an overwhelming wave of AI applications.

Excitement#

This morning, I suddenly discovered that the latest Create GPTs feature of ChatGPT is available.

I was so surprised; I didn't expect it to be so quick. Based on the hints on the page, I thought I would have to wait about another week. So I expressed my feelings:

As a result, just a little over 10 minutes later, my own bot was online. I was so excited that it took me another 10 minutes to remember to add this to the comments section.

In this article, I will share with you the process of creating a GPT bot and its usage effects.

However, before getting to the main topic, I need to provide some background knowledge for some readers—please don't find me verbose; I found that some of my friends still haven't watched the OpenAI DevDay conference.

What are GPTs? To save time, I directly "borrowed" content from Teacher Zhuoke's column on the Dedao App (by the way, highly recommended) and had ChatGPT summarize it for me to share with you:

GPTs are a new feature from OpenAI that acts like an intelligent robot. Three examples were showcased at the conference: CanvaGPT can help you design posters; ZapierGPT can connect and control various apps on your phone, such as checking your schedule, resolving scheduling conflicts, and even helping you explain to your boss on WeChat why you can't attend a meeting; the last example is that you can communicate with ChatGPT in natural language to generate a personalized agent, like a startup consultant. For developers and investors, this new feature is both exciting and concerning, as they need to upgrade their products to compete with OpenAI.

Alright, background introduction complete. Now let's see how to use this feature to construct your own GPT intelligent bot to help you solve important or tricky tasks.

Building#

The process of building a GPT is very easy because it is guided conversationally.

The GPT builder asked me:

Hi! I'm here to help you build a new GPT. You can say something like "Create a creative designer that can generate visual effects for new products" or "Create a software engineer that can help format my code." What would you like to do?

My response was:

Create an article published content query assistant to help me answer questions using already published content.

ChatGPT immediately began to think.

Then it quickly provided the result.

Note the prompt on the right side and several heuristic initial questions provided below. The interface already has a prototype.

Then it continued to ask me:

Great, we're making a content retrieval assistant that will use published articles to answer questions. How about we call this GPT "Article Assistant"? Would that work for you?

Translated to:

Great, we're making a content retrieval assistant that will use published articles to answer questions. We can call this GPT "Article Assistant," does that work for you?

Of course, that was fine with me. I expressed my agreement.

Then, the GPT builder started designing the icon.

After designing, the GPT builder asked me how I felt about the icon.

I thought the icon was not bad, but it was a bit off from the theme. So I suggested it would be better if it related to books or papers.

The GPT builder immediately got to work and modified it to a new icon, as shown above. Then it asked for my opinion.

I said: Very good!

The GPT builder happily asked me for specific details.

Awesome! Now, let's adjust how the Article Assistant works. What type of content should it prioritize when searching for information?

I thought this question was crucial as it could set different priorities and handling methods. I didn't expect this guided conversation to consider the details so thoroughly.

So I replied:

Use the materials I uploaded.

The GPT builder immediately followed up with:

Understood. The Article Assistant will prioritize the content you uploaded for information retrieval. Next, if the assistant cannot find a direct answer in your materials, what should it do? Would you like it to ask clarifying questions or try to provide the best guess based on available information?

Since I was building a query assistant, relying on guessing to answer wouldn't be reliable. So I pointed out:

First, honestly say that there is no answer in the database, and then try to answer using your own retrieval ability + knowledge base, and provide detailed sources, how about that?

The GPT builder's response was:

Perfect, the Article Assistant will notify you when it cannot solve your problem using the files you uploaded and will provide the best answer by utilizing its browsing ability and knowledge base, citing detailed sources. Finally, do you have any preferences on how the assistant should communicate with you? Should it be formal, casual, or somewhere in between?

This question really made me think for a while. Because based on my understanding of ChatGPT's temperament, it can sometimes be quite "rigid." If you say "serious," it tends to respond in a scholarly manner, answering very seriously and even a bit dry; but if you say "lively," it becomes chatty and whimsical. So I thought I would just give it an example:

Quite formal, but not too stiff, like a friendly librarian providing reference consultation to readers.

I was quite satisfied with this analogy, and the GPT builder seemed to understand.

Then it's settled. The Article Assistant will communicate with you in a formal yet friendly tone, like a dedicated librarian ready to assist you with your research. Please feel free to try using the Article Assistant in the playground on the right, and let me know if there are any areas that need further improvement. What would you like to do next?

The remaining tasks didn't require conversation to resolve; I directly clicked on the settings tab.

Scrolling to the bottom, I chose to upload files as the knowledge base for this GPT.

The knowledge base I selected consists of the articles I have published recently (including content from Knowledge Planet and Little Paperboy). A partial article directory is as follows:

I merged these articles into a single knowledge.md file and uploaded it.

Then, I very wisely canceled the default setting below to ensure data privacy.

At this point, our setup phase was complete. Then, I clicked the button in the upper right corner.

When creating a new GPT bot, you can choose from three visibility options:

Creator only
Users with the link
Publicly published

As a test, I initially chose the visibility option to include only myself. Then, the GPT builder automatically exited the settings page and directly entered my own GPTs user interaction interface.

Testing#

I was overly excited and nervously asked Article Assistant (the GPT bot I just built) my first question:

What drawing applications are mentioned in the materials I uploaded? Please elaborate and display the corresponding primary headings for each example.

Article Assistant is thinking...

Soon, it provided the result:

The final result is as follows:

I manually reviewed the answer, and it was accurate, which made me very happy. I tried a few other questions in different directions.

In the morning, I happened to have a group meeting, so I demonstrated it to the students on the spot.

In my database, where is GPT-4 mentioned?

This is Article Assistant's response:

As Article Assistant outputted, my students and I excitedly verified and discussed in real-time. The answers indeed came from the articles I had previously published.

Next, I decided to try invoking the built-in multimodal feature of Article Assistant.

Please use a picture to vividly and realistically depict the content of my database.

Surprisingly, Article Assistant actually drew it.

Let me zoom in; do you think it is suitable for describing my creative themes?

However, this time my students seemed less excited than before and were a bit worried. Qingwei frowned:

Teacher, what will people do in the future?

This question didn't sit well with me. To ease everyone's mood, I let them propose a question to test on the spot.

So Xiaoxiao suggested:

Teacher, ask him what predictions and judgments he made in the articles that later became a reality?

I thought this was a good question. It could compare the knowledge base information with the latest external query results.

So I asked Article Assistant:

Query my database for any potential developments I thought had prospects in the past that have now become a reality?

This is Article Assistant's response:

In your database, there are no specific instances directly mentioning potential developments that were believed to have prospects in the past and have now become a reality. You may need to refer directly to the original materials to determine this information. If you need help identifying specific developments or trends, please provide more context or guidance.

The students looked quite disappointed. I said it depends on whether you can improve the question when facing setbacks. Since the overall answer from Article Assistant was not complete, we could help it break it down. This is one of the techniques of prompt engineering.

So I first asked:

Query my database for any forecasts and judgments, listing them one by one.

Article Assistant did not disappoint us this time.

Then, I continued to ask:

Please query the latest materials to see if the above judgments have become a reality. Provide corresponding examples.

To say that Article Assistant's answers were complete, I wouldn't say so. Because the examples it provided only confirmed the part about "creating job opportunities" from the previous question's answer. However, over the course of two questions, its responses did address the students' initial question. I hope it can improve in the future regarding automatic question breakdown and comprehensive summarization.

Limitations#

In addition, there are still many limitations and shortcomings with GPTs at present.

For example, I posed the following question:

What themes appear most frequently in my database? Please list them and visualize the frequency of occurrence.

It didn't display everything on one page, so I asked it to continue:

Here you can clearly see the problem.

First, there must be an issue with the frequency statistics of the theme words. Because I can't possibly mention important theme words only once. Some of the content above doesn't even count as important theme words. There should be a flaw in the statistical method here.

Additionally, there is still an issue with the Chinese display in the drawing. The Chinese parts of the themes all turned into boxes on the coordinate axes of the graph. I have detailed introduced these issues when I demonstrated the Code Interpreter to you, so I won't elaborate further here.

Summary#

In this article, I introduced you to the latest release of the GPT builder from OpenAI DevDay and its usage effects. By uploading my published articles as a knowledge base, I really constructed my own "knowledge avatar" in just a few minutes and several rounds of natural language dialogue, which can help myself, students, and readers answer questions. I am still immersed in excitement. Although it still has many limitations and problems, I know this is just the beginning. Once the developer's imagination gate is opened, we will inevitably face an overwhelming wave of applications.

In the past few days, I have seen examples of GPTs shared by friends. Among them, what struck me the most was Liqing's sentiment after using ChatGPT to create a competitor, BibiGPT. What kind of directional adjustments do developers need to make? There has already been much discussion in the online space within just a few days, and I don't want to elaborate. I just hope to share Liqing's sentiment to let you experience the intensity of the impact.

Wishing you a pleasant experience in creating AI applications using natural language!

If you have patiently read this far, we must be fated. I invite you to test my Article Assistant together. The link is here.

However, I must give a heads-up:

First, since this is just a test version, I do not guarantee the accuracy of the answers provided by this bot assistant;

Second, it may be "recalled" at any time in the future. If you happen to miss the first opportunity to use it and cannot open the link later, please don't rush to call me a "fraud." 😂

If you find this article useful, please recharge.

If this article might help your friends, please forward it to them.

Feel free to follow my column "Research Tools" to receive timely updates.

Click this link to join the Minority membership for an instant 10% discount! Gain access to exclusive member content, member podcasts, and customized member perks. Open your mind in more fields and directions, find new interests, and gain insights into the present with Minority.

Experience Scroll: Introduction to Academic Paper Writing

Bringing you into academic writing with over twenty years of research experience.

Leo的恒河沙