Summary To address the slow response of traditional well control methods under dynamic geological conditions, their high computational cost, and the tendencies of general-purpose large language models (LLMs) to produce “hallucinations” and encounter knowledge-update difficulties in specialized well control applications, we developed a few-shot retrieval-augmented generation (RAG) system. The system introduces two key innovations: First, it employs low-rank adaptation (LoRA) for deep domain-adaptive fine-tuning of a general embedding model, combined with a hybrid data strategy using synthetic and expert-annotated data, overcoming the limitations of general retrieval models in the well control domain—particularly their low precision and the scarcity of labeled data—and achieving 100% Hitᵣate@5 on the specialized well control retrieval test set. Second, to resolve the issue of basic RAG frameworks often failing to effectively constrain LLM outputs, resulting in noncompliant content, a structured prompt template integrated with industry-standard knowledge bases was designed, which significantly improved the faithfulness and compliance of generated responses, with faithfulness increasing by 13. 3 percentage points. Validated in 40 real industrial well control scenarios, the system achieved a 98% well shutdown procedure completeness score corresponding to a plan completeness (PC) score of 4. 90/5 in expert evaluation and a 95% kill fluid density design compliance rate (KFD-CR; consistent with KFD-CR of 0. 95), delivering an intelligent decision-support solution for high-risk well control operations.
Guo et al. (Wed,) studied this question.