ArcRAG: Agentic Compositional Retrieval over Knowledge Graphs

TL;DR

ArcRAG enables retrieval-augmented generation over knowledge graphs without requiring retrieval strategies to be tailored to specific graph construction methods. It dynamically constructs a query-local interface at inference time and plans multi-hop queries in Cypher, maintaining effectiveness across different graph types. The approach shows that retrieval can remain effective even when the underlying graph structure changes, as long as the interface is adapted at query time.

Why It Matters

This work advances the robustness of knowledge graph-based question answering by decoupling retrieval strategies from graph construction, enabling systems to work across diverse data sources without retraining or redesign. It supports more flexible and scalable deployment in real-world settings where graph schemas vary.

Key Findings

• ArcRAG maintains strong performance across both schema-guided and open-information-extraction graph constructions without retraining or configuration changes.
• Query-time interface induction enables effective multi-hop reasoning even when relation names and property keys differ between graph types.
• While construction method affects absolute accuracy, the core retrieval strategy remains effective across different graph structures, demonstrating cross-construction portability.