Abstract: Based on large language models, an intelligent generation technology for process mineralogy was proposed to address the challenges of low efficiency, insufficient standardization, and high professional requirements in the manual compilation of traditional process mineralogy analysis reports. Built on the Dify platform, the method integrates the Qwen2.5−VL−32B−Instruct large language model to construct a report generation agent encompassing modules for data extraction, knowledge retrieval, prompt−driven generation, and format conversion. For core sections of the report—including mineral composition, liberation degree, intergrowth degree, particle size distribution, elemental occurrence state, and analysis of mineralogical factors affecting beneficiation indicators—the agent employs a differentiated structured prompt system that defines model roles, output format specifications, and logical rules. Additionally, a specialized process mineralogy knowledge base is introduced to enhance the professionalism and accuracy of the generated content. Finally, the LLM−as−a−Judge (Large Language Model as a Judge) method is adopted to conduct multidimensional quantitative evaluations of the generated reports across four dimensions: data accuracy, content completeness, analysis reasonableness, and text standardization. The results demonstrate that the generated process mineralogy reports exhibit clear logical descriptions across all sections, with particularly high professional depth in chapters concerning particle size distribution, liberation degree, and mineralogical factors affecting beneficiation indices.