A dual-stage LLM–KG framework for question answering on oil and gas pipeline topologies

This paper proposes a dual-stage framework for question answering over oil and gas pipeline topologies by coupling a lightweight topology-first knowledge graph with large language models. The framework first performs schema-aligned structural parsing and generates Cypher queries for execution in Neo4j. When direct structured querying cannot provide a reliable answer, it extracts a bounded five-hop local subgraph and performs LLM-based reasoning within this restricted context. By doing so, the method retains the precision of graph querying while reducing the noise and context burden associated with full-graph prompting. Experiments on a representative natural-gas transmission network with 100-, 200-, and 500-node graph instances show that the proposed approach achieves higher end-to-end F1-scores than full-graph and fixed-hop baselines while remaining feasible at larger graph scales. Further analysis indicates that bounded subgraph retrieval offers a better balance between coverage and compactness, although multi-hop questions remain more challenging than one-hop questions. These findings suggest that combining structured graph querying with localized graph-grounded reasoning provides a practical solution for industrial topology question answering.