面向钻井地质设计知识问答的检索增强与生成优化方法

Retrieval-Augmented and Generation Optimized Method for Question Answering in Drilling Geological Design

摘要: 在企业知识问答场景下，大模型支撑系统经常会存在知识缺失、泛化回复、实体关系提取能力弱以及可追溯性不足等问题，因此，本文提出一种面向油田钻井地质设计领域的检索增强与生成优化协同方法。该方法将Ict-Retriever作为检索前端，通过企业文档切片与向量索引实现外部知识召回。同时构建钻井地质设计专用指令微调数据集ICT-DrillGeoDesign，该数据集涵盖地层岩性、压力预测、钻井液设计、井身结构等核心领域知识。除此之外，对ChatGLM3-6B模型进行QLoRA微调，进而得到面向钻井地质设计业务的大模型IctLM-Drill。最后，后续研究将围绕钻井地质设计规范、油气层特征、钻井工程参数以及安全预警标准等核心文档构建领域知识图谱，让系统能够完成地质设计问答、实体关系抽取以及设计参数推荐等任务。实验结果表明，IctLM-Drill在原始ChatGLM3-6B基础上，BLEU、PPL、Distinct指标上分别达到0.458、6.82、9.15。联动Ict-Retriever后，端到端模型指标进一步提升至0.658、6.42、16.87。该方法有效提升了钻井地质设计场景下问答的准确性、专业性和实用性，为油田钻井工程设计提供了智能化支撑。

Abstract: Large language model-based question-answering systems deployed in enterprise environments often face several practical challenges, including incomplete domain knowledge, generic responses, weak capability in extracting entity relations, and limited answer traceability. To address these issues, this paper proposes a retrieval-augmented and generation-enhanced framework for the drilling geological design domain in oilfields. The framework incorporates Ict-Retriever as the retrieval module. Enterprise documents are segmented into manageable passages and encoded into a vector index to support efficient retrieval of relevant external knowledge. In parallel, a domain-specific instruction-tuning dataset, named ICT-DrillGeoDesign, is constructed for drilling geological design tasks. The dataset covers key technical topics, including formation lithology, pressure prediction, drilling fluid design, and wellbore structure design. Based on this dataset, ChatGLM3-6B is fine-tuned using QLoRA to obtain IctLM-Drill, a domain-adapted language model for drilling geological design. In addition, a domain knowledge graph will be developed from core technical materials in the future, including drilling geological design specifications, reservoir characteristics, drilling engineering parameters, and safety warning standards. The knowledge graph enables the system to support multiple downstream tasks, such as domain question answering, entity and relation extraction, and drilling design parameter recommendation. Experimental results show that IctLM-Drill achieves BLEU, PPL, and Distinct scores of 0.458, 6.82, and 9.15, respectively, outperforming the original ChatGLM3-6B model. After integrating Ict-Retriever, the end-to-end system further improves the corresponding results to 0.658, 6.42, and 16.87. These results demonstrate that the proposed framework can effectively improve the accuracy, domain relevance, and practical usefulness of question answering in drilling geological design scenarios, providing intelligent support for oilfield drilling engineering design.