BifrostRAG: Bridging Dual Knowledge Graphs for Multi-Hop Question Answering in Construction Safety
This addresses the problem of automated compliance checking in construction safety for engineers and regulators, representing a domain-specific advancement with competitive gains.
The paper tackles multi-hop question answering for construction safety compliance by introducing BifrostRAG, a dual-graph system that models linguistic relationships and document structure, achieving 92.8% precision, 85.5% recall, and 87.3% F1 score, significantly outperforming baseline methods.
Information retrieval and question answering from safety regulations are essential for automated construction compliance checking but are hindered by the linguistic and structural complexity of regulatory text. Many compliance-related queries are multi-hop, requiring synthesis of information across interlinked clauses. This poses a challenge for traditional retrieval-augmented generation (RAG) systems. To overcome this, we introduce BifrostRAG: a dual-graph RAG-integrated system that explicitly models both linguistic relationships (via an Entity Network Graph) and document structure (via a Document Navigator Graph). This architecture powers a hybrid retrieval mechanism that combines graph traversal with vector-based semantic search, enabling large language models to reason over both the meaning and the structure of the text. Evaluation on a multi-hop question dataset shows that BifrostRAG achieves 92.8 percent precision, 85.5 percent recall, and an F1 score of 87.3 percent. These results significantly outperform vector-only and graph-only RAG baselines that represent current leading approaches. Error analysis further highlights the comparative advantages of our hybrid method over single-modality RAGs. These findings establish BifrostRAG as a robust knowledge engine for LLM-driven compliance checking. Its dual-graph, hybrid retrieval mechanism offers a transferable blueprint for navigating complex technical documents across knowledge-intensive engineering domains.