Building Trustworthy Graph-Agentic RAG for Social Good: Architectures, Failure Propagation, and Assurance by Construction

Summary

This survey examines how graph-agentic retrieval-augmented generation (RAG) — systems combining graph-structured evidence with controllers whose run-time observations change later retrieval, verification, or tool actions — should be designed and evaluated for social-good deployments. The authors organize literature from January 2020 to August 2026 along five mechanism dimensions (graph substrate, lifecycle, agent function, coordination, authority boundary; Table 2), define graph-agentic via a two-condition operational test, and model risks as a five-stage evidence-to-action failure chain (Figure 1) with three propagation mechanisms: amplification, relabeling, and observability loss. The central contribution is an assurance-by-construction blueprint: five interface contracts (evidence, retrieval, reasoning, delegation and capability, outcome), each formalized as input–policy–artifact–validation–failure-response (Table 4), enforcing five invariants (provenance, temporal validity, authorization, uncertainty, recoverability). A hypothetical public-benefits assistant (Table 5) illustrates instantiation, and an evaluation agenda (Table 6) prescribes propagation metrics — origin detection, propagation depth, containment, recovery completeness — plus mandatory comparison against the simplest adequate non-graph, non-agentic baseline. Results are purely conceptual: no experiments, benchmarks, or deployments are reported. Acknowledged limitations (Section 8) include single-reviewer screening without inter-rater agreement, bias toward English indexed literature, and contracts that remain a literature-grounded synthesis rather than a prospectively validated standard; the overhead cost of the required artifacts is never quantified.