Skip to main content
Back to timeline
arXivSource publication:

ArcRAG decouples graph retrieval from how the graph was built: one fixed configuration spans two graph families and still reaches 71.6 F1 on open-vocabulary graphs

Synopsis

Graph retrieval systems are usually designed together with their own graph construction, and whether they survive a change of construction is rarely tested directly. ArcRAG induces a query-local graph interface at inference time from the entities a question names, then plans multi-hop traversals against it and executes and repairs them as Cypher. On MuSiQue, HotpotQA and TriviaQA, one fixed configuration runs over both a schema-guided graph and an open-information-extraction graph, reaching 38.0 EM, 65.5 EM and 71.6 F1 on the open-vocabulary graphs, while the schema-guided graphs score higher by 6.0 EM, 9.0 EM and 16.3 F1.

Source-provided article image: ArcRAG: Agentic Compositional Retrieval over Knowledge Graphs
arXiv

Interpretation

One retrieval configuration runs unchanged over two structurally different knowledge graphs, so retrieval need not be co-designed with graph construction. Earlier graph retrieval methods each presuppose one construction product: personalized PageRank assumes a dense synonym-linked entity graph, GraphRAG global search assumes a hierarchical community partition, LightRAG assumes an entity-and-relation keyword index, and PathRAG selects paths by a heuristic over a precomputed path index. Each of MuSiQue, HotpotQA and TriviaQA is built as a schema-guided graph and as an open-information-extraction graph, with the retriever configuration fixed across them; the open-vocabulary graphs yield 38.0 EM, 65.5 EM and 71.6 F1.

The system first establishes which relations and paths actually exist around the entities a question names, then plans and executes queries against them. A Direct-Cypher agent reads only the introspected schema, which says how the graph is stored, not what lies near a question. On a stratified 100-question sample, adding exploration beats direct Cypher generation at all 18 retry and pass settings, by 15 to 39 points on the schema-guided graphs and 12 to 32 on the open-vocabulary graphs.

Planning helps depending on the graph type rather than uniformly across constructions. On the schema-guided graphs, planning beats exploration plus generation when the Generator has a single retry and gives the best accuracy at every retry-pass budget up to 9. On the open-vocabulary graphs, planning is no better than exploration plus generation at small budgets and leads by 3 to 5 points at larger budgets, within sampling noise at this sample size.

Perspective

The results rest on the encyclopedic QA corpora MuSiQue, HotpotQA and TriviaQA, each built as one schema-guided graph and one open-information-extraction graph that differ in structure, relation vocabulary and provenance. The experiments use encyclopedic QA and do not cover finance, law, healthcare or compliance settings that motivate traceable retrieval.

The two graph families differ in both structure and provenance, so the accuracy gap between them reflects their combined effect and is not separated. How much noise or variation query-time canonicalization absorbs is not measured. Comparisons with prior GraphRAG systems use their best reported figures under different protocols and answer models rather than reruns in a shared setup. Each configuration runs once, and the answer model samples at temperature 1.0 and judges its own evidence sufficiency, so results carry unquantified run-to-run variance. Construction still shapes absolute accuracy, so co-design of retrieval and construction is not ruled out.