My Eureka moment on Chat GPT
When Chat GPT was GPT 3
When GPT 3 was released, my data scientist son told me to try it out. It was still GPT 3, not Chat GPT yet and had curiosity only in select geek community. Since it is my part of my job profile is to keep our product line upto date, I never ignore any such suggestion. I did some basic exploring and completely blown away with it. The accuracy and reasoning of being able to understand context was very close humans. I felt we can use GPT3 davinci model to replace some of NLP in-house model we are using, which were giving around 75 to 80% accuracy. During further exploring I came across a road block. Training GPT3 on large data set of the enterprise was a major bottleneck. One challenge was Proteus Vision is a comprehensive ERP and has large number of master information. Another bigger challenge our organisation Proteus Technologies is a pure cloud based service organisation. Each of our customer have specific master information. Unless the AI is able to understand this master information, it would be impossible for it to interpret the context. For any AI capability to be used in business application, the information available in the internet is not good enough. I gave up the idea of using ChatGPT capability in our application and continued focus on improving our in house models.
Chat GPT was promising
In the meantime GPT3 came as Chat GPT and stormed the world. It quickly came with another version, which had option to train it with your own data. Followed by options to write custom ChatGPT plugins. There is a large amount of information available on how to train Chat GPT on your data, I won’t discuss them here. I am sure they have their own use-case and it is an excellent way of using the platform. However in my experience they require lot of efforts and still expensive. Here you have to understand I am in India and our clients are SME, so yes it becomes expensive. Our customers are not that comfortable in sending their data out to train Chat GPT. Though when I requested OpenAI to not use our data for any training purpose, filling an opt out form, they instantly gave a confirmation.
The Eureka moment
I tried to see if there is any other use cases besides NLP in chat GPT. Even I tired asking this question to mighty Chat GPT3 itself. It gave quite a few options, but all of them are NLP expressed in different ways. I did continue to use Chat GPT to write articles and make presentation contents for marketing activities. No, this one is not written by Chat GPT :). After almost a year I encountered this Eureka moment, I was watching a b-grade web series, I heard a dialogue “If you can’t find a solution, let the solution find you”. It did not make sense on the series, but it made sense to me; In the context of using Chat GPT. If Chat GPT can’t understand data of Vision ERP, let Vision ERP understand Chat GPT data. I did some SQL generation experiments on Chat GPT console by asking Chat GPT to create SQL from business question.
It gave accurate result with incorrect table and column names(naturally as Chat GPT has no idea about my table and column names) . But it always gave me the same wrong table and column names. To use this capability in application, all we had to do is, create a simple mapping from Chat GPT to Vision ERP terminology mapping. In 3 weeks we went live with our 1st Chat GPT powered question answer feature. We did a lot more development, such a schema designer to create logical table with pre-joined tables to further increase accuracy and reduce complexity of join, in natural language to SQL question answering feature.
Same way we used Chat GPT for intent and entity detection in our Chat Bot Vision Assistant. We allow the Chat GPT to give the name of intent and entity in its own nomenclature and we use a dictionary feature within our product to translate the terminology.
The journey continues
There was no looking back, with the concept of vanilla Chat GPT prompt with custom components to understand and interpret the response, we have over 10 major Chat GPT specific AI features in our Proteus Vision ERP. Check out some more example of AI feature here on our product website Proteus Vision ERP . Many of them are powered with Chat GPT using this concept. It also worked like a game changer for us in the market place, as our product looked visibly more AI powered compared to competition.
We started off by using davinci model, but later migrated to get-turbo-3.5 this reduced or API cost by 10 times. Just that we had rephrase the prompt and got the exact same result. Though for different use cases we are using different model of Chat GPT.