Answering Natural Language Questions with OpenAI's GPT in the Petroleum Industry

Abstract

Abstract This work documents two experiments that make use of OpenAI's ChatGPT and GPT-4 for question answering in the petroleum industry. First, we describe PetroQA, a prototype tool that can answer natural language questions. It uses PetroWiki content to inform ChatGPT about facts specific to this industry. We are able to convince ChatGPT to avoid hallucinations and cite its sources. We asked nearly 200 SPE members to volunteer to test PetroQA and discuss results from that test. Second, we are developing and testing a tool, known as GraphQA, that allows users to ask questions and receive answers from a large graph knowledge base consisting of facts and relations between concepts such as wells, fields, basins, formations, geography, geologic age, rock type, operators, and more. A knowledge base like this is difficult for users to explore, so we use GPT-4 to automatically generate accurate graph queries from their natural language questions. We explore several novel techniques for prompting GPT-4 to produce the right queries and have developed an advanced caching mechanism to reduce interactions with the cloud model, thus reducing time to answer and cost.