Of course, I don’t just talk and write about ChatGPT, I also experiment with it myself.

My first experiments

I started with simple chats using the well-known chatbot ChatGPT from OpenAI. First, I sent general requests to ChatGPT, such as “Explain to me how Large Language Models work!”. I then generated complex texts with the help of ChatGPT. So I was able to learn a lot about correct prompting myself. I have recorded some of the results here.

The ChatGPT experiment at Helvetia

Helvetia Insurance ‘s ChatGPT experiment started around February 2023 and I was allowed to accompany the project on behalf of the Lucerne University of Applied Sciences and Arts. We have shown that it is possible to combine ChatGPT’s language model with your own company website content to offer a completely new chatbot experience. During this phase, we were able to show that this type of chatbot can be implemented much faster than classic intent-based chats. Overall, everyone involved in the project was able to learn a lot about GPT3.5 and chatbots during this phase. For example, we learned how content should be presented on the website and which content is best suited for website chatbots.

My own ChatGPT chatbot experiment

Now the time has finally come, I can say that I have my “own” ChatGPT chatbot. I use the latest offer from Enterprisebot. The Enterprisebot platform uses my website content as a knowledge base for “my own ChatGPT Bot” . The chatbot combines the ChatGPT3.5 language model with the entire knowledge of my website.

My own ChatGPT chatbot experiment: Procedure

Whereas with my other bots I had to define all the intents, training sentences and answers in advance, this time I simply handed over my website to the Enterprisebot team and they incorporated my content into their platform. The platform has been using the OpenAI language models for a few months now. A few tests and design adjustments followed and then the chatbot went online. The tests have already shown that the bot is not yet able to answer all questions, so we have added a default message.

My own ChatGPT chatbot experiment: Learnings

Magic – the bot is still in the experimental phase, but we were able to show that a simple FAQ bot can be created with very little effort.

I also noticed that some of the information on my website is incomplete. In this case, the chatbot cannot find a suitable answer. The success of the chatbot is therefore not solely down to the language model, but also to the appropriate content.

I will certainly have to do my own homework over the next few months and add to the content on my website.

My own ChatGPT chatbot experiment: Data protection

My new chatbot uses both the Enterprisebot chatbot platform and ChatGPT from OpenAI. The Enterprisebot platform only collects data that is required for the provision of the chatbot service. ChatGPT may collect and process personal data such as name, address, telephone number or e-mail address if these are entered in the chat. It is possible that some of this personal data that could identify users, such as name, address or telephone number, is stored, but not systematically. It is therefore best not to enter any personal information in your conversation in order to be completely safe. The data will never be passed on to third parties unless this is required by law or necessary to fulfill our contractual obligations.

Enterprisebot uses the ChatGPT service from Azure. The data is stored in a cloud within the EU.

By using my new chatbot experiment, you acknowledge and recognize that the interaction is processed with technology from and in the United States. The provisions of OpenAI at www.openai.com/policies apply .

My own ChatGPT chatbot experiment: Join in

Of course, you can also test the chatbot yourself. Simply click on the chat bubble at the bottom right of this page.

Further experiments

Since mid-April, “my previous chatbot provider¨ from Germany has also been offering the integration of ChatGPT in a chatbot. So far, this bot has been a classic intent-based chatbot that I had to “feed” with intents, training sentences and answers. With the help of the new ChatGPT connection, the chatbot can access the entire ChatGPT knowledge, not just individual web pages. “My” Chatbot can now be used as a kind of digital assistant that answers questions like the well-known ChatGPT. I have reconfigured my chatbot so that it offers to forward the chat to ChatGPT if the user request is not understood. If the user agrees, the bot forwards the request to ChatGPT and then returns the response from ChatGPT to the user. This bot is integrated on the homepage of my website. Just click here.

Please note that ChatGPT can only answer general inquiries. If a user has a very specific question about me that they can’t answer based on my “trained” intents, it could also become more difficult for ChatGPT.

Conclusion

ChatGPT and the APIs of OpenAI are already very powerful today. Earlier bots had to be trained with a lot of manual effort. Today, the combination of a good company website and the GPT language model can be used to create a functioning FAQ bot in a short space of time.

It is important that the content source, i.e. the website, is comprehensive and has enough content as a knowledge base. Furthermore, the company’s own tonality and target group-specific communication should not be completely disregarded. This information can be supplemented with the help of fine-tuning.

ChatGPT and other LLMs will certainly shape the world of chatbots, set new standards and make new applications of chatbots in communication possible. It’s best to start gathering your own experience early on. My ChatGPT and LLM Canvas can be a helpful tool here.