The chatbot was built using data from the American multinational financial service Wells Fargo. The chatbot is named WALLY and is prompted to be clever, creative and friendly.
The first step in creating the chatbot was collecting the necessary data for fine tuning. Since I could not find a dataset online that fit the specific requirements, I performed web scraping to collect the data. Product descriptions were collected directly from the Wells Fargo website and customer testimonials were gathered from www.sitejabber.com.
The testimonials collected from sitejabber were varied which I broke into 3 different categories, “bad reviews”, “good reviews” and “mixed reviews”. Once collecting the data, I compiled them into an excel sheet.
Once the data was collected I used jupyter notebook to organize the data and convert it into json format. I then used openai data preparation tools to convert the json file into a jsonl format as required by openai to perform the finetuning operation. The jsonl file was then used to train the davinci model. The davinci model was chosen because it is the most powerful and flexible model.
On completion of successfully creating a fine-tuned model. I then created a basic flask application and webhook to connect to Dialogflow. Dialogflow was used to create the conversational interface for the chatbot. The flask application was then hosted on replit and published.
The chatbot can be accessed via the link https://replit.com/@moemoola01. The project is called WFchatbot. The code to create the webhook can also be viewed here.
- Click into the WFchatbot tile
- At the top of the project make sure the repl is running
- You can now use the link to access the chatbot interface https://WFchatbot.moemoola01.repl.co
