基于GraphRAG的洗涤流程个性化推荐应用研究OA
Personalized recommendation of laundry cycles:an application study based on GraphRAG
随着消费者对衣物精细化护理需求的日益增长,传统洗衣机固化的洗涤程序和参数已难以满足复杂多变的个性化场景.为在高效去渍与精致护衣间取得平衡,提出一种基于知识图谱增强RAG的混合推理架构.该方法构建涵盖衣物材质和颜色、污渍类型、洗护工艺等多维知识的洗护领域知识图谱,并行执行图谱查询与RAG检索以获取结构化因果知识和上下文信息,经知识融合后由大语言模型生成精确洗涤参数及自然语言建议.实验表明,该架构在参数准确率和安全性上显著优于基线方法.该混合智能范式有效解决了在复杂决策场景下AI的可信度与可解释性问题,推动智能家电进入"专家决策"生成式AI时代.
With the rising consumer demand for sophisticated garment care,the pre-set programs and parameters of traditional washing machines are struggling to meet the needs of complex and diverse personalized scenarios.To strike a balance between efficient stain removal and delicate fabric care,a hybrid reasoning architecture leveraging Knowledge Graph-enhanced Retrieval-Augmented Generation(RAG)was proposed.The method involves constructing a domain-specific knowledge graph for laundry care,which encompasses multi-dimensional knowledge such as fabric materials and colors,stain types,and washing-care techniques.The architecture performs parallel queries on the knowledge graph and retrieval via the RAG framework to acquire both structured causal knowledge and unstructured contextual information.Following a knowledge fusion stage,a Large Language Model(LLM)generates precise washing parameters along with natural language recommendations.Experimental results demonstrate that the proposed architecture significantly outperforms baseline methods in both parameter accuracy and safety.The hybrid intelligence paradigm effectively addresses the challenges of trustworthiness and interpretability for AI in complex decision-making scenarios,advancing smart appliances into an era of"expert-level decision"generative AI.
李凯龙;张斌;刘运涛;胡娟;王昱坤;席明轩;朱建雄
小米科技(武汉)有限公司 湖北 武汉 430070小米科技(武汉)有限公司 湖北 武汉 430070小米科技(武汉)有限公司 湖北 武汉 430070小米科技(武汉)有限公司 湖北 武汉 430070小米科技(武汉)有限公司 湖北 武汉 430070东南大学 江苏 南京 211189东南大学 江苏 南京 211189
信息技术与安全科学
智能洗衣机知识图谱检索增强生成大语言模型个性化洗涤
Smart washing machineKnowledge graphRetrieval-augmented generation(RAG)Large language model(LLM)Personalized washing
《家电科技》 2026 (4)
44-49,53,7
评论