Transforming Historical Drilling Data into Strategic Insights with Large Language Models

This paper explores the integration of Large Language Models (LLMs), such as DeepSeek, with unstructured data repositories in upstream petroleum operations. The goal is to develop an LLM-driven assistant that processes historical drilling documents to support strategic well planning, with a focus on risk mitigation and operational efficiency. Our approach involves combining a Retrieval-Augmented Generation (RAG) pipeline with an LLM to interpret unstructured daily drilling reports. Embedding of historical content is performed using a domain-optimized model (BGE-Mining/v1) and stored in a vector database for semantic retrieval. The assistant retrieves relevant historical narratives based on queries about specific drilling challenges or spatial intervals (e.g., nearby wells or formations). Upon receiving a user query, the system returns structured insights such as prior incidents, intervals of concern, and recommended corrective measures. Preliminary tests using a carefully designed corpus demonstrated that the LLM-based assistant can reliably identify recurring drilling risks, such as stuck pipe events, abnormal flow signatures, pressure-related anomalies, and improper mud weight reductions. These insights were extracted from simulated field data constructed to replicate authentic operational reporting styles, including realistic technical language, domain-specific abbreviations, and event sequences. The system effectively interpreted both explicit incidents and subtle early-warning signs, such as pit gain trends and SICP fluctuations. Responses to user queries were generated within minutes and achieved high scores in expert evaluations for contextual accuracy and technical clarity. Furthermore, the architecture incorporated spatial filtering mechanisms, enabling the assistant to simulate retrieval of offset well data based on user-defined coordinates or formation intervals. This spatial logic was essential for testing the assistant's ability to prioritize relevant records within a given geographical scope, further validating its application to location-aware risk assessment tasks in pre-drill planning. This framework highlights the potential of LLMs to transform pre-drill planning workflows by surfacing actionable lessons from complex, text-heavy drilling archives. The proposed system demonstrates a novel and practical application of AI in upstream planning, enabling engineers to proactively manage risks using historical knowledge through a natural language interface. Future enhancements may further increase accuracy and reliability by incorporating modular components that independently handle retrieval, validation, and reasoning.