Source-linked AI summary
Building Trustworthy Graph-Agentic RAG for Social Good: Architectures, Failure Propagation, and Assurance by Construction
Vijay Bommireddy, Raviteja Bommireddy
TL;DR
Graph-agentic RAG can support tasks requiring relational evidence, but coupled failures can carry defects toward consequential outcomes. This paper synthesizes architectures and risks, proposes five interface contracts, and derives an evaluation agenda; it presents these as an evaluable reference architecture rather than a guarantee.
Problem
Graph-agentic RAG systems need design and evaluation methods that preserve traceability and control operating authority from source to consequential outcome.
Method
The paper organizes literature by graph substrate, lifecycle, agent function, coordination, and authority, then develops an evidence-to-action failure model and five interface contracts.
Results
The paper produces a public-benefit information design and evaluation agenda spanning graph assertions, trajectories, grounding, coordination, abstention, human review, and recourse.
Takeaways & Limitations
The framework makes provenance, temporal validity, authorization, uncertainty, and recoverability explicit at system boundaries.
Takeaways & Limitations
The synthesis is not a prospectively validated standard, deployed service, legal determination, certification procedure, or safety guarantee.
Abstract
from arXiv · showhide
Graph-agentic retrieval-augmented generation combines structured evidence with adaptive controllers that can plan retrieval, traverse relations, verify intermediate claims, delegate subtasks, and use tools. This combination is useful when answers depend on relations across documents, entities, time, or institutions, but it also creates coupled failure paths: a defect in graph construction can become retrieved evidence, alter later control decisions, and propagate toward a consequential outcome. We examine how such systems should be designed and evaluated for social-good settings in which freshness, authorization, traceability, oversight, and recourse matter alongside answer quality. We organize the literature by graph substrate, graph lifecycle, agent function, coordination pattern, and authority boundary, and distinguish graph-based retrieval from observation-dependent graph control. We then synthesize reported risks as an evidence-to-action failure chain and propose an assurance-by-construction blueprint comprising five interface contracts for evidence, retrieval, reasoning, capability and delegation, and outcome. These contracts make provenance, temporal validity, authorization, uncertainty, and recoverability explicit at system boundaries. An illustrative public-benefit information design shows how the framework constrains graph structure, permissions, abstention, and operating authority. Finally, we derive an evaluation agenda spanning graph assertions, trajectories, claims, coordination, and outcomes.
1 Introduction
Graph-agentic RAG addresses relational evidence problems but introduces coupled paths from graph defects to consequential outcomes. The paper proposes an assurance-oriented framework that makes evidence, authority, and recovery explicit.
- Motivation: Graph defects can be retrieved, alter later controller decisions, pass through handoffs, and appear as confident answers.The paper frames retrieval poisoning, graph poisoning, memory failures, and indirect prompt injection as propagation risks.
- Research question: The central question is how to keep relational evidence traceable and operating authority controllable from source to consequential outcome.
- Contributions: The paper synthesizes graph substrates, lifecycles, adaptive control, coordination, and authority boundaries.
- Contributions: Its evidence-to-action failure model explains how defects propagate across graph construction, retrieval, reasoning, coordination, and oversight.
- Contributions: Five interface contracts cover evidence, retrieval, reasoning, delegation, and outcome, assigning preventive, detective, and recovery responsibilities.
- Contributions: A public-benefit information design and reporting agenda address graph quality, trajectories, grounding, coordination, abstention, human review, and recourse.
- Scope: The blueprint is a literature-grounded design synthesis, not a certification procedure or safety guarantee.
2 Scope and evidence discipline
The review uses a targeted literature-search and screening process, while explicitly bounding its evidence base and synthesis claims.
- Evidence scope: The review covers work from January 2020 through 29 August 2026, with earlier foundational studies added through relevant citation chaining.
- Search strategy: Searches combine RAG, graph, knowledge-graph, agent, planning, and orchestration terms with provenance, freshness, robustness, privacy, fairness, oversight, and contestability.
- Selection: Inclusion requires a graph or retrieval substrate, generation, and enough control detail to identify adaptive decisions, alongside directly relevant assurance work.
3 From Graph Retrieval to Graph-Agentic Systems
GraphRAG and agentic RAG encompass diverse mechanisms, so the paper defines graph-agentic systems by material graph evidence plus observation-dependent control. It recommends matching complexity and authority to task requirements.
- System scope: GraphRAG systems vary from entity and community summaries to existing-graph subgraphs, learned paths, and query-time relation construction.
- Operational boundary: A system is graph-agentic when graph structures materially support the answer and execution observations can change later operations.
- Design dimensions: The design space separates evidence substrate, agent decisions, coordination, and authority, with authority distinguishing information support from world-changing action.
- Failure interaction: Graph construction and agent control interact because uncertain extracted edges can guide traversal while limited verification context hides source qualifications.
- Assurance: Assurance should attach to interfaces rather than only to model components.
- Design choice: The minimal-design guide adds graphs for material relational structure, agentic control for observation-dependent steps, and both only when traces, stopping rules, and permissions are enforceable.
- Related systems: Representative systems illustrate distinct graph and adaptive-control mechanisms, but their results are not pooled because tasks and protocols differ.
- Design question: The key design question is which relations and runtime decisions matter, and what evidence and authority must cross each interface.
4 The Evidence-to-Action Failure Chain
Graph-agentic systems create a coupled five-stage failure chain in which defects can originate upstream, alter later retrieval or control, and propagate toward consequential outcomes. Evaluation must therefore track not only failures, but their origins, propagation, interception, and reversibility.
- Five-stage failure chain: The five-stage chain spans source and graph integrity, retrieval and traversal, reasoning and grounding, coordination and action, and human governance and outcomes.Stages are non-exclusive and may repeat through feedback, memory updates, delegation, or new tool calls.
- Source and graph integrity: Graph defects such as stale, unauthorized, incomplete, or malicious assertions can become retrieved evidence and redirect later generation.Entity resolution, relation extraction, and persistent edges introduce transformation risks before retrieval begins.
- Retrieval and traversal: Retrieval errors are path-dependent because entry-node choices, biased traversal, missed exceptions, temporal mismatches, and stopping rules influence subsequent queries.Adaptive retrieval makes the first returned edge part of the later control state.
- Reasoning and grounding: Relevant evidence may still fail to entail a claim when versions conflict, unsupported relations are inferred, or uncertainty is expressed as certainty.A graph path requires semantic meaning, source support, and time and access context; citation correctness and answer quality remain separate measures.
- Coordination and outcomes: Coordination and action can amplify errors through indirect instructions, poisoned memory, ambiguous handoffs, delegated messages, and tool calls.A technically faithful answer can still cause harm when warnings, routing, review authority, or correction channels are inadequate.
- Evaluation implications: Trustworthy evaluation should measure where failures originate, whether they are intercepted before the next interface, how far they propagate, and whether effects can be reversed.Amplification, relabeling, and observability loss make end-to-end accuracy an incomplete safety signal.
5 Assurance by Construction
Assurance by construction places explicit validation obligations at interfaces where evidence and authority move between stages. Five contracts preserve key invariants, constrain operating authority, and define fail-closed responses when artifacts cannot be checked.
- Interface contracts: Five interface contracts require evidence, claims, instructions, and action requests to cross boundaries with enough context for validation.Each contract emits an observable artifact, validates it before release, and specifies a response to failure.
- Contract structure: The contract tuple records accepted input, governing policy, observable artifact, pre-release validation, and failure response.This boundary formulation lets downstream stages reject incomplete artifacts and trace failures to the first violated contract.
- Contract structure: Schemas and thresholds remain domain-specific, but evidence and authority should not cross a boundary without an observable basis for checking them.The proposed table associates each interface with an artifact, validation, and failure response.
- Invariants: The contracts preserve provenance, version and temporal scope, access conditions, inspectable reasoning context, and explicit authority boundaries.Retrieved instructions, model outputs, and inter-agent messages cannot create new tool or data permissions; outcome handling also requires accountable review and correction.
- Operating authority: Architecture and operating authority should be chosen separately, ranging from exploratory information to mandatory review or no automated final action.A simpler design is preferred when evidence is bounded, deterministic relations suffice, adaptive control is unnecessary, or required traces cannot be maintained.
6 Illustrative design: Public-Benefit Information
The public-benefit example applies the contracts to a temporally and relationally complex information task. It uses bounded graph-agentic retrieval while keeping permissions constrained, traces inspectable, and final determinations with authorized humans.
- Task and architecture: The example addresses a resident’s question about programs, rule versions, exceptions, jurisdiction, dates, and missing information after a move and income change.A single semantic search may retrieve a plausible general rule while missing an exception or using the wrong date.
- Task and architecture: A typed, versioned graph supports the task because relational and temporal structure matters, while observation-dependent steps may trigger questions, exception inspection, or version comparison.The design uses bounded graph-agentic retrieval but does not justify multi-agent architecture or execution tools.
- Trace and permissions: The complete trace retains source, issuer, jurisdiction, validity interval, version, retrieval path, filters, stopping reason, and clarifying questions.Unsupported or unofficial evidence is quarantined, and missing information triggers clarification rather than guessing.
- Trace and permissions: The contracts map material claims to supporting evidence, keep conflicting rule versions visible, restrict approved sources, and prevent retrieved instructions from expanding permissions.Final determinations are routed to authorized caseworkers while correction and contestation channels remain available.
- Design consequence: The blueprint changes the objective from answering a multi-hop query to controlling how evidence and authority move through the system.Graph assertions become source-backed and time-bounded, retrieval remains inspectable, least privilege is enforced, and consequential decisions stay outside the system’s authority.
7 Evaluation and Reporting Agenda
The paper proposes evaluating graph-agentic RAG beyond answer accuracy by measuring evidence, trajectories, authority, propagation, recovery, and human effects. It calls for stress tests and baselines that reveal whether added mechanisms justify their failure surface and operating cost.
- Evaluation should cover graph assertions, execution trajectories, delegated authority, and human effects alongside retrieval relevance, answer faithfulness, and citation quality.
- Origin detection, propagation depth, containment, and recovery completeness measure whether defects are recognized, contained, and corrected across system interfaces.
- Stress tests should vary graph, evidence, authority, handoff, and permission conditions, with expected responses such as seeking evidence, abstaining, denying action, or escalating.
- Every evaluation should use matched non-agentic and, where appropriate, non-graph baselines while reporting quality, latency, cost, trace completeness, and intervention burden.
- Research priorities include assertion provenance, time validity, access labels, auditable trajectories, permission-aware handoffs, and outcome studies co-designed with affected communities.
8 Limitations
The review’s evidence base and synthesis are constrained by limited screening independence, language and indexing choices, changing preprint status, and heterogeneous studies. Its contracts and case are literature-grounded proposals requiring validation with practitioners and affected communities.
- The review used one primary reviewer with a consistency audit, searched mainly English-language indexed literature, and included fast-moving preprints whose status may change.
- Heterogeneous tasks, graphs, models, and metrics prevent defensible score pooling, so the paper compares mechanisms rather than ranking architectures.
- The contracts and worked case are a literature-grounded synthesis, not a prospectively validated standard, deployed service, or legal determination.
- Their utility and burden require testing with domain practitioners and affected communities.
9 Conclusion
Graph-agentic RAG is appropriate only when relational evidence and observation-dependent control are material, not as a universal upgrade. The proposed failure chain and five contracts make evidence and authority traceable through outcomes, while remaining an evaluable reference architecture rather than a guarantee.
- Graph-agentic RAG is useful when relational evidence and observation-dependent control are both material, but it is not a universal upgrade.
- In consequential settings, evidence quality and authority must survive the complete path from source to outcome.
- The failure chain and five contracts make that path observable through versioned, replayable, and recoverable artifacts.
- The framework is an evaluable reference architecture rather than a guarantee, requiring graph audits, trajectory stress tests, permission-aware agent evaluation, and outcome studies.
- The safest design is the least complex one whose evidence is traceable, authority bounded, and effects reversible.