Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SCOUT-RAG: Scalable and Cost-Efficient Unifying Traversal for Agentic Graph-RAG over Distributed Domains

Longkun Li, Yuanben Zou, Jinghan Wu, Yuqing Wen, Jing Li, Hangwei Qian, Ivor Tsang

TL;DR

SCOUT-RAG addresses the challenge of scalable, cost-efficient reasoning over distributed, access-controlled knowledge by formulating retrieval as a sequential decision problem across multiple domains. It introduces a three-stage, four-agent architecture that estimates domain relevance, seeds domain-scoped retrieval, and iteratively refines answers under explicit budget constraints, without centralized graph access. The approach combines training-free relevance estimation, adaptive local/global traversal, and quality-guided expansion to balance retrieval cost and grounding fidelity. Experimental results on 89 multi-domain queries across 45 countries show SCOUT-RAG achieves near-parallel performance to centralized baselines like DRIFT while delivering substantial reductions in tokens and latency, demonstrating practical scalability for cross-domain knowledge synthesis.

Abstract

Graph-RAG improves LLM reasoning using structured knowledge, yet conventional designs rely on a centralized knowledge graph. In distributed and access-restricted settings (e.g., hospitals or multinational organizations), retrieval must select relevant domains and appropriate traversal depth without global graph visibility or exhaustive querying. To address this challenge, we introduce \textbf{SCOUT-RAG} (\textit{\underline{S}calable and \underline{CO}st-efficient \underline{U}nifying \underline{T}raversal}), a distributed agentic Graph-RAG framework that performs progressive cross-domain retrieval guided by incremental utility goals. SCOUT-RAG employs four cooperative agents that: (i) estimate domain relevance, (ii) decide when to expand retrieval to additional domains, (iii) adapt traversal depth to avoid unnecessary graph exploration, and (iv) synthesize the high-quality answers. The framework is designed to minimize retrieval regret, defined as missing useful domain information, while controlling latency and API cost. Across multi-domain knowledge settings, SCOUT-RAG achieves performance comparable to centralized baselines, including DRIFT and exhaustive domain traversal, while substantially reducing cross-domain calls, total tokens processed, and latency.

SCOUT-RAG: Scalable and Cost-Efficient Unifying Traversal for Agentic Graph-RAG over Distributed Domains

TL;DR

SCOUT-RAG addresses the challenge of scalable, cost-efficient reasoning over distributed, access-controlled knowledge by formulating retrieval as a sequential decision problem across multiple domains. It introduces a three-stage, four-agent architecture that estimates domain relevance, seeds domain-scoped retrieval, and iteratively refines answers under explicit budget constraints, without centralized graph access. The approach combines training-free relevance estimation, adaptive local/global traversal, and quality-guided expansion to balance retrieval cost and grounding fidelity. Experimental results on 89 multi-domain queries across 45 countries show SCOUT-RAG achieves near-parallel performance to centralized baselines like DRIFT while delivering substantial reductions in tokens and latency, demonstrating practical scalability for cross-domain knowledge synthesis.

Abstract

Graph-RAG improves LLM reasoning using structured knowledge, yet conventional designs rely on a centralized knowledge graph. In distributed and access-restricted settings (e.g., hospitals or multinational organizations), retrieval must select relevant domains and appropriate traversal depth without global graph visibility or exhaustive querying. To address this challenge, we introduce \textbf{SCOUT-RAG} (\textit{\underline{S}calable and \underline{CO}st-efficient \underline{U}nifying \underline{T}raversal}), a distributed agentic Graph-RAG framework that performs progressive cross-domain retrieval guided by incremental utility goals. SCOUT-RAG employs four cooperative agents that: (i) estimate domain relevance, (ii) decide when to expand retrieval to additional domains, (iii) adapt traversal depth to avoid unnecessary graph exploration, and (iv) synthesize the high-quality answers. The framework is designed to minimize retrieval regret, defined as missing useful domain information, while controlling latency and API cost. Across multi-domain knowledge settings, SCOUT-RAG achieves performance comparable to centralized baselines, including DRIFT and exhaustive domain traversal, while substantially reducing cross-domain calls, total tokens processed, and latency.
Paper Structure (26 sections, 8 equations, 4 figures, 1 table, 1 algorithm)

This paper contains 26 sections, 8 equations, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Problem Formulation
  5. Stage I. Domain Relevance Assessment
  6. Domain Relevance Assessment Agent (DRAA).
  7. Stage II. Domain-Scoped Seeding
  8. Partial Answer Generation Agent (PAGA).
  9. Overall Answer Synthesis Agent (OASA).
  10. Stage III. Iterative Cross-Domain Refinement
  11. Answer Quality Assessment Agent (AQAA).
  12. Strategy Selection.
  13. Partial Answer Generation Agent (PAGA).
  14. Overall Answer Synthesis Agent (OASA).
  15. Experiments
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Overview of the proposed SCOUT-RAG framework. The framework operates in three stages: (i) domain relevance estimation, (ii) domain-aware retrieval (local vs. global), and (iii) iterative answer refinement via synthesis and generation agents. These agentic components collectively support adaptive, cross-domain reasoning without centralized graph access.
  • Figure 2: Illustration of Stage III, where the system evaluates and improves answers through selective retrieval and synthesis. Example query: "What are the main differences in dietary habits among countries such as Japan, France, and the United States?".
  • Figure 3: Time-performance monitoring and comparison.
  • Figure 4: Case study on Italy's "Made in Italy" certification.