Cross-institutional knowledge discovery in the context of stringent regulatory environments such as the GDPR and HIPAA requires making a difficult choice between centralizing document repositories in vector databases at the expense of institutional sovereignty or employing broadcast federated search that suffers from O(N) network overhead, divulges query intentions, and becomes more susceptible to prompt injection attacks. SentinelRAG solves the problem by decoupling knowledge discovery orchestration from local document repository management. Sovereign institutions keep all proprietary documents, dense embeddings, and BM25 indices behind their firewalls but reveal opaque Ed25519-signed capability triples on a central Neo4j overlay graph (dIKG). On receiving queries, the orchestrator queries up to k ≤ 2 eligible nodes in the dIKG using paths through the graph and requests excerpts from them, which must be accompanied by cryptographic proof envelopes generated via the SHA-256 Merkle proof system. The platform was benchmarked on BEIR SciFact (50 queries, 3 partitions), an enterprise technical document set (120 queries, 3 partitions), and a STRIDE adversarial test suite of 140 instances. SentinelRAG constrains cross-node communications to 1.2–2.0 edges per query, achieves 0.965 citation accuracy along with 0.982 synthetic consistency, and defends against all attempts at forgery, injection, and tampering attacks within 15–32 ms of local overhead. The 0.44 recall@10 drop when compared to the fully pooled dataset is thus formally calculated as the cost of partitioning.

Retrieval-augmented generation,decentralized systems,edge data sovereignty,hybrid search,cryptographic provenance,attribute-based access control,knowledge-graph routing