Generative AI Driven Hybrid Vision-Language and Property Graph Pipeline for Reliable Information Retrieval
Book Chapter
Etar, A, Valle, S, Soni, J et al. (2027). Generative AI Driven Hybrid Vision-Language and Property Graph Pipeline for Reliable Information Retrieval
. 2967 CCIS 221-232. 10.1007/978-3-032-28454-9_15
Etar, A, Valle, S, Soni, J et al. (2027). Generative AI Driven Hybrid Vision-Language and Property Graph Pipeline for Reliable Information Retrieval
. 2967 CCIS 221-232. 10.1007/978-3-032-28454-9_15
We investigate the limitations of Retrieval-Augmented Generation (RAG) in retrieving all relevant content from complex, safety-critical document and show how a property graph-based approach overcomes these challenges. Conventional RAG pipelines rely on semantic similarity to retrieve partial text snippets, which often fail to capture deeply nested or image-encoded information and can hallucinate unsupported content. In contrast, our hybrid ingestion uses PyMuPDF for heuristic text and table extraction and Mistral Small 3.2 as the vision-language model for page-level layout recovery, yielding a near-lossless PDF parse that preserves structure and tables. We then persist the reconstructed hierarchy as a Neo4j property graph, enabling deterministic, fully cited retrieval in which queries return exact text excerpts from the appropriate nodes with high recall and precision. Compared to a semantic RAG baseline on SMR FSER Q&A tasks, graph search delivers complete, citation-backed answers while RAG often omits critical details. Our results surface real world obstacles to exhaustive retrieval and highlight the advantages of explicit structure and tool choice.