SCOUT-RAG: Distributed Graph-RAG Framework

• SCOUT-RAG is a distributed system that integrates retrieval-augmented generation with unsupervised, agent-based control to optimize answer quality while minimizing cost and latency.
• It employs a three-stage process involving domain relevance assessment, seeding via partial answer generation, and iterative cross-domain refinement to balance depth and breadth effectively.
• Performance evaluations show that SCOUT-RAG rivals centralized methods by significantly reducing token usage and latency, making it ideal for privacy-sensitive applications like hospitals and multinational corporations.

SCOUT-RAG is a distributed, agentic Graph-RAG (Retrieval-Augmented Generation using structured knowledge graphs) framework that enables scalable and cost-efficient retrieval over siloed or access-restricted knowledge domains. It is designed for environments where centralized knowledge graph construction is infeasible due to privacy, regulation, or ownership constraints, such as in hospitals or multinational corporations. SCOUT-RAG performs progressive, utility-guided cross-domain traversal, leveraging a closed-loop of cooperative LLM-based agents to optimize answer quality under strict cost and latency constraints while minimizing retrieval regret, defined as the utility lost by not retrieving from useful domains. The framework achieves performance approaching centralized and fully exhaustive decentralized Graph-RAG methods at a fraction of the cross-domain API and computational cost, and introduces a suite of algorithmic strategies, metrics, and agentic controls for privacy-aware multi-domain retrieval (Li et al., 9 Feb 2026).

1. Motivation and Problem Setting

Retrieval-augmented generation (RAG) approaches augment LLMs with information retrieved from structured knowledge sources. Graph-RAG, in particular, improves multi-hop and entity-relation reasoning by integrating LLMs with centralized knowledge graphs. However, in distributed real-world settings, consolidation into a global graph is prevented by data silos and access restrictions. Each domain $\mathcal D_i$ exposes only a local graph API, typically at a domain-specific cost $c_i$ (encompassing token, API, or latency expenses).

Three core challenges arise in this distributed, access-restricted Graph-RAG scenario:

• Partial Observability: No global graph is visible, only per-domain $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$ accessed through isolated APIs.
• Cost–Quality Trade-off: Full cross-domain traversal and exhaustive querying are typically cost-prohibitive or too slow for practical use.
• Absence of Supervised Domain Routing: Training examples mapping queries to relevant domains are rarely available, especially at system cold start.

SCOUT-RAG addresses these by providing unsupervised, dynamic, and sequential domain selection and traversal, budgeting API calls under user-set constraints ($\mathcal C_{\max}$ cost, $T_{\max}$ time), and aiming to minimize retrieval regret while maximizing answer quality (Li et al., 9 Feb 2026).

2. Framework Architecture and Core Algorithm

SCOUT-RAG operates in three sequential stages, coordinated by four specialized LLM-based agents:

1. Domain Relevance Assessment:

The Domain Relevance Assessment Agent (DRAA) computes, for each domain $i$, three signals: - $s_i^{\mathrm{sim}} = \mathrm{Sim}(q, \mathcal D_i)$: query–domain embedding cosine similarity. - $$s_i^{\mathrm{rich}} = |\mathcal R_i|/\max_j|\mathcal R_j|$$: normalized report/data count. - $$s_i^{\mathrm{hist}} = (1/|\mathcal H_i|)\sum_{h\in\mathcal H_i}Q(h)$$: historical average answer quality. DRAA assigns each domain a relevance tier: HIGH, MODERATE, POTENTIAL, or IRRELEVANT.

2. Domain-Scoped Seeding: The Partial Answer Generation Agent (PAGA) retrieves globally from HIGH and locally from MODERATE domains, with POTENTIAL domains reserved. Partial answers $\mathcal A_i$ are generated and synthesized into an initial seed answer $c_i$0 by the Overall Answer Synthesis Agent (OASA):

$c_i$1

$c_i$2

3. Iterative Cross-Domain Refinement: The Answer Quality Assessment Agent (AQAA) evaluates answer completeness $c_i$3, diversity $c_i$4, and knowledge gaps $c_i$5, proposing follow-up queries $c_i$6. A Strategy Selector decides among Depth (further exploration of HIGH domains), Breadth (engaging POTENTIAL domains), Hybrid, or Stop. The process stops when $c_i$7, the budget ($c_i$8 or $c_i$9) is exhausted, or answer quality converges.

A “best-track” answer $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$0 is maintained to prevent performance deterioration in late iterations.

3. Mathematical Formulation and Optimization Criteria

The framework’s retrieval-augmented optimization objective is: $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$1 where $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$2 denotes domain-specific retrieval policy, $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$3 the corresponding retrieval operator, and $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$4 the answer synthesis function.

Retrieval regret is informally defined as the difference in total grounding utility between the optimal subset $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$5 and the policy-selected subset $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$6: $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$7

The iterative agentic refinement is governed by: $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$8 with strategy selection thresholds: $$\mathcal G_i=(\mathcal V_i,\mathcal E_i,\mathbf X_i)$$9

Targeted retrieval at each refinement step: $\mathcal C_{\max}$0

4. Cooperative Agent Roles and Control Loop

SCOUT-RAG coordinates four specialized agents:

• Domain Relevance Estimator (DRAA): Assigns domains to relevance tiers using similarities, data-size ratios, and historical answer quality, producing both the discretized score and a rationale.
• Strategy Selector: Dynamically decides when to explore new domains, deepen search within current domains, or terminate, based on metrics $\mathcal C_{\max}$1 and remaining time.
• Traversal Depth Adapter: Orchestrates further multi-hop retrievals within HIGH domains when completeness is low.
• Answer Synthesizer (OASA): Aggregates partial answers, enforces consistency, and archives the answer with the highest observed quality.

All agents operate without supervised domain labels, enabling cold-start and privacy-sensitive deployments. The framework follows a closed control loop, executing DRAA → PAGA+OASA → AQAA → Strategy Selector, with up to $\mathcal C_{\max}$2 refinement iterations, where $\mathcal C_{\max}$3 is the average duration per loop.

5. Experimental Protocol and Results

Experiments utilized 45 independent “country” knowledge graphs from Wikipedia, each with 9–77 community reports, and 100 multi-domain natural-language queries (89 answered by all systems). Queries spanned single-domain and up to very large (40-domain) regimes.

Baseline methods included centralized GraphRAG (both local/entity and global/summary searches), centralized DRIFT-c (one global plus two local refinement rounds), and fully decentralized DRIFT-dec (DRIFT applied independently to each domain).

Key quantitative results (averaged over 89 queries):

Method	Overall Quality	Time (s)	Tokens
Centralized GraphRAG-local	53	34.4	11,223
Centralized GraphRAG-global	49	45.9	640,574
Centralized DRIFT-c	63	231.9	693,731
Decentralized DRIFT-dec	85	414.9	879,911
SCOUT-RAG	56	75.3	159,169

SCOUT-RAG achieved equivalent overall quality to centralized DRIFT-c (56 vs. 63) while reducing token usage by 77% and latency by 67%. Against the fully decentralized DRIFT-dec, SCOUT-RAG operated 81.9% faster (75 s vs. 415 s) and consumed 81.9% fewer tokens, at a 29-point deficit in overall quality (56 vs. 85). Notably, SCOUT-RAG outperformed both centralized local and global GraphRAG on diversity (60 vs. 55/50), attributed to its tiered, quality-guided domain activation.

Additional analysis demonstrated that answer quality rises substantially within the first 120 seconds and saturates by 180 seconds, indicating diminishing returns for longer retrieval cycles. Case analyses, such as for "Made in Italy" certification, illustrated rapid identification of relevant domains, efficient seed generation, and effective refinement.

6. Cost–Quality Trade-offs, Deployment, and Real-World Considerations

SCOUT-RAG is positioned for scenarios where exhaustive, cross-domain retrieval is prohibitively expensive or slow. Its training-free, signal-driven domain ranking and selective traversal afford rapid, cost-effective approximation to centralized or fully distributed Graph-RAG methods. Cost–quality trade-offs are tunable via strategy thresholds (e.g., completeness or time values in Eq. 5). Lowering completeness thresholds induces greater breadth/diversity, potentially at the expense of core accuracy, while higher thresholds focus on depth and completeness.

Because relevance estimation is unsupervised and employs only semantic similarity, data size, and historical quality, SCOUT-RAG requires no prior domain–query training, improving cold-start viability. As historical performance $\mathcal C_{\max}$4 accrues, domain assignment becomes increasingly precise.

The framework is readily adaptable: domains only need to expose a compatible PAGA interface and domain embedder, with lightweight LLM prompts for AQAA/OASA. This facilitates deployment across varied enterprise, governmental, or federated environments.

7. Summary and Frontier Implications

SCOUT-RAG introduces an agentic, privacy-aware, and cost-controlled approach to distributed Graph-RAG, operationalizing a sequential, utility- and quality-driven retrieval strategy over siloed API-accessible graphs. It balances local versus global retrieval, depth versus breadth, and explicit utility–cost trade-offs, delivering performance near centralized baselines at a fraction of retrieval cost and latency. It is the first such framework to enable cold-start deployment, adaptive multi-agent refinement, and practical estimation of retrieval regret in distributed knowledge settings (Li et al., 9 Feb 2026).