Leveraging ChatGPT and Tidio for building a Chatbot

One of the most best aspects of working with AI tools like ChatGPT is the ability to process and analyze datasets to extract meaningful insights. This week’s project involved analyzing a dataset from specific Shopify stores, focusing on product details, links to products, product category, (T-shirt, hat, tote bag, etc), Blank style #, ordering method for blanks (stocked versus order) and decoration methods/locations. The goal was to have ChatGPT extract insights, generate questions, and organize the information into a structured format that would be helpful for the customers.

My initial results, before adding Chat GPT’s CSV, were not exactly what I expected. The agent wasn’t accurately pulling information from the non-ChatGPT dataset. Answers were too random, I quickly realized that getting Tidio to use the dataset effectively required more work. First, I had to translate the raw data into a structured Q&A format. Thats where ChatGPT stepped in.

Using ChatGPT for Data Analysis and Insight Extraction

My first step was to provide ChatGPT with a dataset that contained information about the store’s products. I had the information in excel and I exported it as a CSV. Keep in mind this doesn’t always work. In this case, everything was weblinks or text, and not too complex, so it works. Understanding your data and how an AI will interact with it is key. I can test this by providing the data to Chat GPT and interacting, if it gets confused easily, I need to consider another format. Next, I needed the AI to process this information, generate questions and answers an e-commerce customer might have, and present everything in a CSV file format. The questions a human might have took a little direction. downloading the CSV and then asking the AI to consider … X, Y, Z but ultimately testing is where you will start to understand any gaps.

ChatGPT analyzed the dataset, filtering for specific attributes like decoration methods, product categories, and collections within my store. It wasn’t just about pulling raw data; ChatGPT could generate direct links to products, ensuring the answers were informative and actionable.

What I found particularly useful was how ChatGPT refined its approach based on my feedback. Initially, the answers weren’t as comprehensive as I needed. However, after a few tweaks, the AI expanded the list of questions and pulled more relevant insights.

Translating Data for Tidio Q&A: The Next Step in the Process

Once the data was formatted into a ChatGPT Q&A CSV, I moved on to the next phase; integrating it with Tidio, a customer support chatbot I am testing for Shopify stores. I uploaded the CSV into Tidio’s knowledge base for the Lyro agent and decided to exclude data pulled directly from other web stores. I had hoped that Tidio could use this data for a more comprehensive response (as it is an option in the knowledge to include websites) but the results were too inconsistent. Web stores were built for human interaction, not AI-driven responses. Expecting the AI to interpret that kind of data became too complex, the responses were too wordy, so I only focused on using the dataset from Chatgpt Q&A CSV. I’m still unsure whether this was the right call, but I will have to do more testing to know. More is not always better.

Why Tidio?

With Tidio, I will be able to automate responses and improve the customer service experience. The tool costs $29/month. I’m looking forward to the next step, which is Beta testing. This phase will allow me to iron out any issues, refine the Q&A format further, and test how well the AI handles customer inquiries in real-world situations. If it passes Beta, I will update this blog with a link to the stores. ChatGPT is $20/month.

The beauty of this project lies in AI’s flexibility. With ChatGPT processing my data and Tidio answering questions based on that data, I can streamline customer interactions, reduce response time, and provide better support. As I move into Beta testing, I’ll continue refining the data input and the AI’s ability to understand and respond accurately. I did discover in Beta already that any questions involving the word “Gun” for example End Gun Violence T-Shirt ​in Tidio, no matter how the user asks the question, the response is “Unfortunately, what you inquired about exceeds my current capabilities . Would you like support on another matter, perhaps? .” In ChatGPT the answer was correct, but the URL did not work, even though it worked in the provided dataset. This reflects the guardrails on AI – trying to protect users from harmful content. As you can see it is generic.

This project took about 4 hours, moving forward updates will be 15 minutes or less.

Here is a video of the action: