A Human–AI Interactive Agent Collaboration Framework and Its Application to Dam Seepage Safety Assessment

This study proposes a human–computer interactive agent collaboration framework for reservoir dam seepage safety assessment. This is to address the large volume of safety assessment reports, inconsistent textual descriptions, low efficiency of manual statistics, and difficulty in maintaining consistent classification criteria. Using a large language model as the core for semantic understanding, the framework decomposes the report review process into subtasks including data preprocessing, knowledge retrieval, semantic classification, result feedback, and manual review. Structured collaboration among multiple agents is achieved through an “entity location-safety hazard” knowledge primary key, semantic rule constraints, and the Model Context Protocol (MCP). While preserving expert control, the system transforms original texts into aggregable and traceable risk statistics, forming a closed-loop mechanism of “machine-enabled efficiency, human oversight, and knowledge accumulation”. Application validation was conducted using more than 2000 reservoir safety assessment reports, from which 2724 valid severe-risk records were extracted. The results show that severe risks were mainly concentrated in key locations, including dam bodies, culverts, spillways, dam foundations, downstream dams, drainage facilities, and abutments. The main risk types were leakage, abnormal seepage behavior, construction quality problems, noncompliant seepage control, structural damage, and noncompliant drainage facilities. The findings indicate that the proposed framework can improve the efficiency of text processing for seepage safety assessments and the consistency of risk statistics, thereby providing intelligent support for reservoir dam hazard screening and scientific decision-making.