ArcRAG
A graph-based retrieval-augmented generation (GraphRAG) method that dynamically generates a query-local graph interface at inference time, enabling cross-construction portability of the retrieval pipeline across graph construction methods.
Timeline (last 90 days)
ArcRAG is a retrieval method that dynamically generates a query-local graph interface at inference time, enabling the retriever to work without co-design with a specific graph construction method.
ArcRAG identifies entities in the question, collects reachable node labels, relation descriptors and attribute keys, and forms a local interface specific to the current query only.
ArcRAG compiles the planned multi-hop traversal path into Cypher query statements for execution and has repair capability after execution failures.
On the three corpora MuSiQue, HotpotQA and TriviaQA, ArcRAG cross-validated by constructing a schema-guided graph and an open information extraction graph, respectively.
Experiments show that a construction-agnostic retriever remains effective after the graph construction method is changed, i.e., retrieval does not need to be co-designed with graph construction.
The graph construction method (schema-guided vs. open information extraction) still affects the absolute accuracy of the final answers.
Ablation experiments under different retry and pass budgets reveal that ArcRAG's planning, execution and repair stages each contribute to the final performance.
ArcRAG's findings support decoupling the retrieval module from the graph construction module in GraphRAG systems, thereby reducing system integration complexity.
ArcRAG is currently presented as a paper; specific implementation details and code availability should be referred to in the original text.
All Facts (9)
ArcRAG is a retrieval method that dynamically generates a query-local graph interface at inference time, enabling the retriever to work without co-design with a specific graph construction method.
50%UnverifiedArcRAG identifies entities in the question, collects reachable node labels, relation descriptors and attribute keys, and forms a local interface specific to the current query only.
50%UnverifiedArcRAG compiles the planned multi-hop traversal path into Cypher query statements for execution and has repair capability after execution failures.
50%UnverifiedOn the three corpora MuSiQue, HotpotQA and TriviaQA, ArcRAG cross-validated by constructing a schema-guided graph and an open information extraction graph, respectively.
50%UnverifiedExperiments show that a construction-agnostic retriever remains effective after the graph construction method is changed, i.e., retrieval does not need to be co-designed with graph construction.
50%UnverifiedThe graph construction method (schema-guided vs. open information extraction) still affects the absolute accuracy of the final answers.
50%UnverifiedAblation experiments under different retry and pass budgets reveal that ArcRAG's planning, execution and repair stages each contribute to the final performance.
50%UnverifiedArcRAG's findings support decoupling the retrieval module from the graph construction module in GraphRAG systems, thereby reducing system integration complexity.
50%UnverifiedArcRAG is currently presented as a paper; specific implementation details and code availability should be referred to in the original text.
50%