Optimizing Semantic Search in Industrial Knowledge Retrieval: A Novel SHAP-Based Attention Mask Modification Approach

Knowledge management in industrial settings faces significant challenges due to information overload and knowledge loss. This investigation examines a RetrievalAugmented Generation (RAG) solution for quality reports at Siemens Electronics Works, where engineers currently spend significant time searching through <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$4,500+$</tex> technical documents, resulting in annual costs of approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathrm{C}\!\!\!\!{\scriptstyle{{}^=}\,} 200,000$</tex>. We developed an optimized data preprocessing pipeline for semantic search that combines generalizable components (duplicate removal, HTML cleaning, chunking strategies) with novel domain-specific enhancements. Our key innovation is a systematic approach integrating Shapley Additive exPlanations (SHAP) with attention mask modification (AMM) to identify and downweight tokens that disproportionately influence search results. This method, combined with LLM-based contextual enhancement of domain-specific abbreviations and semantic chunking, achieved a 21.1 % improvement in retrieval accuracy compared to raw data processing, as measured by Mean Reciprocal Rank (MRR). The resulting RAG application operates locally without internet access, ensuring data security while significantly reducing document retrieval time. Our methodology provides a framework for optimizing semantic search in specialized technical domains without requiring extensive model retraining.