How we made our first custom GPT agent

In this blog post, I'll share my experience creating a custom GPT and the lessons I learned along the way.

My goal is to develop a custom GPT trained with public twoday.lt data to answer questions for interested parties, whether about our career opportunities, modernization service, or anything else.

Custom GPT model creation

Let's roll up our sleeves and get started. When opening the create new GPT page, you'll find two tabs on the left for refining and configuring your agent and a playground on the right to test how this custom GPT works. 

You can train this custom GPT by providing instructions. For instance, I instructed GPT to explore the twoday.lt website and acquire comprehensive knowledge about twoday. Since these custom GPTs are in their early stages, they may make errors, but I found that asking them to repeat the action often resolves the issue.

Now that we have basic information about twoday, we can conveniently configure this agent. By using the GPT builder, you can initiate the configuration process with a simple command: 'Please update the name, description, instructions, and conversation starters based on the information you've learned.' In just a few seconds, you'll have a prefilled configuration that can be fine-tuned manually or through multiple rounds of the GPT builder. For example you ask to add an instruction to cut off access to internet and provide answers only from GPT’s knowledge base.

Additionally, you can expand GPT's knowledge base by uploading your documents. Keep in mind that there's currently a limit of 10 documents. If you have a large amount of files, consider merging them.

I've made some updates to the name, description, and instructions to better align with the specific context of twoday and avoid generic references to the tech world. Let's put our pretrained GPT to the test and see if it can accurately answer a question about twoday.

Is it really that simple?

As can be seen it is quite easy to generate your GPT and do the initial training. However, such GPT will only answer a handful of questions correctly as we don't have control of what it now knows. By giving this twoday.lt URL I can only assume that it took all the parts from it as we can’t see what actually the GPT learned from the website, but surely it had its own way to approach this data and it may not be the best one.  

To ensure control over your data, it's advisable to structure it yourself and upload well-organized information to GPT. This way, you can be confident that the model has learned the information accurately or, at the very least, has been given the best opportunity to do so 🙂

Keep an eye out for the soon to be available twoday Lithuania Agent in the GPT store.