Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Divide by Question, Conquer by Agent: SPLIT-RAG with Question-Driven Graph Partitioning

Ruiyi Yang, Hao Xue, Imran Razzak, Shirui Pan, Hakim Hacid, Flora D. Salim

TL;DR

SPLIT-RAG tackles the efficiency-accuracy tension in retrieval-augmented KGQA by partitioning a knowledge graph into type-aligned subgraphs guided by training questions, and by distributing retrieval and reasoning across lightweight agents assigned to those subgraphs. A retrieval planner selects a small, relevant set of agents and decomposes incoming questions into sub-tasks, enabling parallel subgraph retrieval, while a conflict-resilient merger cleans and weights evidence before a head LLM generates the final answer. The approach is proven information-preserving and achieves substantial latency reductions, with theoretical guarantees on retrieval complexity and mutual-information preservation. Empirically, SPLIT-RAG attains state-of-the-art performance on WebQSP, CWQ, and MetaQA benchmarks, illustrating robust accuracy improvements with scalable, efficient KG reasoning.

Abstract

Retrieval-Augmented Generation (RAG) systems empower large language models (LLMs) with external knowledge, yet struggle with efficiency-accuracy trade-offs when scaling to large knowledge graphs. Existing approaches often rely on monolithic graph retrieval, incurring unnecessary latency for simple queries and fragmented reasoning for complex multi-hop questions. To address these challenges, this paper propose SPLIT-RAG, a multi-agent RAG framework that addresses these limitations with question-driven semantic graph partitioning and collaborative subgraph retrieval. The innovative framework first create Semantic Partitioning of Linked Information, then use the Type-Specialized knowledge base to achieve Multi-Agent RAG. The attribute-aware graph segmentation manages to divide knowledge graphs into semantically coherent subgraphs, ensuring subgraphs align with different query types, while lightweight LLM agents are assigned to partitioned subgraphs, and only relevant partitions are activated during retrieval, thus reduce search space while enhancing efficiency. Finally, a hierarchical merging module resolves inconsistencies across subgraph-derived answers through logical verifications. Extensive experimental validation demonstrates considerable improvements compared to existing approaches.

Divide by Question, Conquer by Agent: SPLIT-RAG with Question-Driven Graph Partitioning

TL;DR

SPLIT-RAG tackles the efficiency-accuracy tension in retrieval-augmented KGQA by partitioning a knowledge graph into type-aligned subgraphs guided by training questions, and by distributing retrieval and reasoning across lightweight agents assigned to those subgraphs. A retrieval planner selects a small, relevant set of agents and decomposes incoming questions into sub-tasks, enabling parallel subgraph retrieval, while a conflict-resilient merger cleans and weights evidence before a head LLM generates the final answer. The approach is proven information-preserving and achieves substantial latency reductions, with theoretical guarantees on retrieval complexity and mutual-information preservation. Empirically, SPLIT-RAG attains state-of-the-art performance on WebQSP, CWQ, and MetaQA benchmarks, illustrating robust accuracy improvements with scalable, efficient KG reasoning.

Abstract

Retrieval-Augmented Generation (RAG) systems empower large language models (LLMs) with external knowledge, yet struggle with efficiency-accuracy trade-offs when scaling to large knowledge graphs. Existing approaches often rely on monolithic graph retrieval, incurring unnecessary latency for simple queries and fragmented reasoning for complex multi-hop questions. To address these challenges, this paper propose SPLIT-RAG, a multi-agent RAG framework that addresses these limitations with question-driven semantic graph partitioning and collaborative subgraph retrieval. The innovative framework first create Semantic Partitioning of Linked Information, then use the Type-Specialized knowledge base to achieve Multi-Agent RAG. The attribute-aware graph segmentation manages to divide knowledge graphs into semantically coherent subgraphs, ensuring subgraphs align with different query types, while lightweight LLM agents are assigned to partitioned subgraphs, and only relevant partitions are activated during retrieval, thus reduce search space while enhancing efficiency. Finally, a hierarchical merging module resolves inconsistencies across subgraph-derived answers through logical verifications. Extensive experimental validation demonstrates considerable improvements compared to existing approaches.

Paper Structure

This paper contains 39 sections, 3 theorems, 27 equations, 3 figures, 12 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. RAG with LLMs
  4. Knowledge Graph Partition
  5. LLM-based multi-agent systems
  6. Divide by Question, Conquer by Agent: SPLIT-RAG framework
  7. Question-Type Clustering and Knowledge Graph Partitioning
  8. Question preprocessing.
  9. $\mathcal{Q}_{\text{train}}$-guided KG partitioning.
  10. Subgraph--Agent--Question Matching
  11. Matching objective.
  12. Properties
  13. Retrieval Plans Decision for Incoming Questions
  14. Agent selection.
  15. Question decomposition.
  16. ...and 24 more sections

Key Result

Theorem 1

Given the balancing factor $\lambda$ in Algorithm alg:alg_gp, the graph partitioning achieves: where $I(\cdot;\cdot)$ denotes mutual information, $H(\cdot)$ is entropy, $\mathcal{Q}$ is the question distribution, and $\mathcal{G}$ is the subgraph collection.

Figures (3)

  • Figure 1: Complete structure of the SPLIT-RAG framework, with knowledge base preprocessing, retrieval plan decision generation, subgraph partition, multi-agent RAG, and answer generation.
  • Figure 2: Time--Hit comparison with Pareto frontier (frontier around G3--G5--G6).
  • Figure 3: Performance degradation relative to the best H@1. Darker color indicates larger degradation.

Theorems & Definitions (6)

  • Theorem 1: Information-Preserving Partitioning
  • proof
  • Theorem 2: Semantic Interpretability
  • proof
  • Theorem 3: Search Space Reduction
  • proof