Enhancing large language models for fashion smart manufacturing via dynamic collaborative routing-based retrieval reranking

Abstract

Enhancing large language models (LLMs) with external knowledge base retrieval in the fashion manufacturing industry can provide more reliable technical support and decision-making assistance, significantly improving process control and boosting intelligent production efficiency. However, the field of fashion manufacturing involves highly specialized terminology, logically complex technical knowledge, and intricate query tasks. Existing simple query-matching techniques often return a large number of contextually loose and redundant document chunks, severely impacting the model’s understanding and response quality. To address this issue, this article proposes a retrieval optimization framework based on a dynamic capsule routing network with embedded semantic graph (SGDCR), which models semantic relations among multiple retrieved documents by simulating a team collaboration mechanism. Specifically, the framework consists of two steps: filtering and reranking. First, a capsule routing mechanism embedded in a semantic association graph dynamically captures complex contextual relationships among coarse-grained document blocks, learns contribution scores for multiple documents, and filters irrelevant or redundant documents based on ranking. Subsequently, the filtered documents are matched with the query through deep semantic similarity measurement, and the documents are reranked by integrating relevance scores and contribution scores and generating efficient, accurate, and contextually coherent document prompts. Experimental results on publicly available dense open-domain QA datasets and a constructed fashion manufacturing process QA dataset demonstrate the effectiveness and superiority of the proposed method over existing reranking approaches in the fashion manufacturing knowledge QA system.