M.S. Candidate: Şeymanur Özen
Program: Data Informatics
Date: 31.08.2026 / 13:00
Place: A-212

Abstract: Technical documentation for enterprise electronic products is a primary source of  engineering and product knowledge. It spans several formats, such as datasheets and  brochures, each presenting information differently: brochures in narrative text,  datasheets in tables and product-specific numeric specifications. Because the same  product is described at different levels of detail across these document types, locating the required information is challenging, and semantic similarity alone is often insufficient for engineering queries. When a query concerns a specific product, conventional retrieval-augmented generation (RAG) systems do not account for  product-level relationships or attributes, and may retrieve chunks belonging to  different products together, producing misleading answers. 

This thesis proposes a graph-guided RAG architecture for question answering over electronic product documentation. Product relationships and attributes are modelled  in a Neo4j knowledge graph, and queries are filtered through this graph before vector search to restrict retrieval to the matched product scope. A rule-based intent router  directs each query to the appropriate processing flow, and a tiered chunk-gating  mechanism enforces this scope during retrieval. 

The proposed method is compared against a hybrid RAG baseline on a benchmark  dataset prepared as part of this work, using standard retrieval evaluation metrics (F1,  Hit@K, Recall@K) alongside LLM-based judge evaluations. The results show that  the graph-guided approach reduces cross-product contamination by 78%, with  statistically significant improvements in retrieval accuracy and answer quality.  These findings indicate that incorporating product-level relationships and attributes  into RAG improves question answering over technical documentation.