Hybrid Vector–Graph Retrieval-Augmented Generation for Enterprise Resource Planning: Contextual Grounding, Referential Integrity, and Hallucination Mitigation

Abstract

Retrieval-augmented generation (RAG) has become the default mechanism for grounding large language models (LLMs) in private corpora, yet the dominant implementation, dense vector similarity over flat text chunks, is a poor structural match for enterprise resource planning (ERP) data. ERP records are not prose: they are normalised, heavily cross-referenced, temporally versioned tuples whose meaning derives from join paths (purchase requisition → purchase order → goods receipt → supplier invoice → payment) rather than from lexical or semantic proximity. A retriever that scores chunks independently by cosine similarity has no mechanism for guaranteeing that the documents it returns are mutually consistent or that they refer to the same business object, and the resulting evidence sets invite a characteristic failure mode we term referential hallucination: fluent output that correctly cites individual documents but silently fabricates the relationships between them. This paper proposes HVG-RAG, a hybrid vector–graph retrieval architecture for ERP question answering, and makes three contributions. First, we formalise referential integrity as a first-class retrieval-time constraint rather than a post-hoc evaluation criterion, and specify a constraint-satisfaction verifier that operates over an ontology-bound property graph induced from ERP metadata. Second, we introduce an adaptive fusion policy in which the weight assigned to vector versus graph evidence is predicted per query from an estimated reasoning depth, rather than fixed by a global hyperparameter. Third, we attach bitemporal validity intervals to graph edges so that “as-of” queries, pervasive in finance and audit, are answered against the state of record at the relevant time rather than against current master data. We develop the accompanying mathematics, give complexity bounds, and specify a reproducible evaluation protocol with ERP-specific metrics. We evaluate the framework on ERP-Bench-P2P, an executable procure-to-pay benchmark released with this paper: 19,632 documents projected from a normalised schema, a 47,914-edge foreign-key graph, and 1,700 questions stratified by reasoning depth with gold provenance. The results reported here are measured on that benchmark rather than simulated. At a retrieval budget of 100 documents, hybrid dense-sparse retrieval answers 100% of one-hop questions and 0.6% of two-hop questions, because the documents joined by a foreign key share almost no vocabulary; HVG-RAG answers 100%, 91.5%, 99.1% and 97.9% of one- to four-hop questions at a measured p95 latency of 96 ms. Two findings run against our own expectations and we report them as such: five-hop questions requiring selection of the contract revision effective at the transaction date remain largely unsolved (57.1% at k = 100, 0.3% at k = 20), and a graph-only retriever is competitive with the proposed system whenever the query names an identifier, its advantage disappearing only when the anchor entity must first be located by description. End-to-end generation quality is not measured here and remains future work.