The tool allows you to talk to a digital assistant about a selection of the Norwegian Petroleum Directorate’s publicly available status reports, well histories and news articles. You can ask questions about wells, discoveries, geology, prospects and more.
“You can choose between four personalities to assist with your questions”, says Jesse Lord, founder and Lead – Product Strategy at Kadme. Jesse developed the search tool earlier this year, around the time many people became familiar with the ChatGPT service. The tool is based on the same digital language models as the popular chatbot.
“We wanted to launch this quickly so that those working with the applications for this year’s APA Round can test it out. We believe the artificial intelligence assistant has helped them do research, as it understands the context of questions and sits on a lot of valuable information”, explains Petter Dischington, geologist at the NPD.
“We wanted to see if, with few resources, we could make available useful information for internal work processes and for those who use our data..”
“In addition, we wanted to explore how we can use this type of technology internally. We wanted to see if, with few resources, we could make available useful information for internal work processes and for those who use our data, using language models and user-friendly interfaces”, adds NPD geologist Maria Juul.
It is Petter and Maria who have been the driving force behind the initiative at the NPD.
A friendly face
The assistants you meet at the platform on the NPD website, Nina and Neil, are based on the latest language models GPT 3.5 and GPT 4. The language models have been developed by OpenAI and are based on so-called neural networks. It is the tremendous increase in available computing power and quantity that has contributed to such models becoming possible in the past year.
“The tool we have developed at Fabriq is based on OpenAI’s language models, but is specifically designed for the oil and gas industry. The knowledge base for the digital assistants is, as of today, publicly available data from the NPD. Based on the users’ questions, the most relevant data is fed into the language model, which makes sure to provide an understandable answer in a natural language form”, explains Jesse.
Building trust
Many of those who have tried ChatGPT have painfully learned that you cannot fully trust the information that is presented. First, the chatbot does not tell you where it got its information from. Secondly, it just so happens that it gives you information that is not rooted in reality.
ChatGPT is primarily developed to be able to give a well-formulated and relevant answer. Facts come second. “The first feedback we gave Fabriq was that all information presented by the assistant must be able to be documented”, says Petter.
A large part of the work on the development of the tool has actually been to give users the answer they don’t want: “Sorry, but I have no information about…”. “It was absolutely necessary for us to ensure that the assistant does not fabricate an answer”, points out Jesse.
The tool, which is still in a pilot phase, can still be used but will probably be further developed soon.
