Intelligent Knowledge Generation and Evidence-based Research on Accident Investigation Reports Based on DeepSeek

In order to solve the problems that the application of large language models in the field of security engineering is restricted by factors such as limited corpus resources, limited input capacity and data privacy, a localized incident question-answering system combining the DeepSeek large language model and the RAG mechanism is constructed. Realize intelligent parsing and knowledge services for complex texts to assist in security management decisions. Based on the accident reports and laws and regulations released by the government emergency management system, a semantic feature corpus is constructed, and technologies such as PaddleOCR, LayoutLMv3, and YOLOv8 are integrated to complete the document structure reconstruction and semantic modeling. The system covers four stages: document parsing, semantic alignment, knowledge base construction and hybrid retrieval, and has the capabilities of causal chain extraction, regulation matching and semantic mapping. The actual measurement shows that, compared with the Deepseek-r1:32b model without using the RAG mechanism, the automatic score of system question answering has increased by 7.7%, and the manual score has increased by 17.6%. It performs better in terms of response speed and stability. Despite being limited by the scale of local model parameters and the knowledge update mechanism, it still shows good adaptability and practicability, providing technical support for intelligent accident handling and having the potential to be extended to emergency response and early warning scenarios.