StatementGraphRAG: Cross-Model Graph Retrieval

• StatementGraphRAG is a retrieval-augmented generation paradigm that unifies RDF and property graphs using a directed acyclic graph abstraction.
• It employs bidirectional mappings and cross-model querying with SPARQL and Gremlin to support multi-domain, context-sensitive question answering.
• The system, demonstrated via the 1G Playground, enables stack-independent data exchange and seamless semantic integration across heterogeneous sources.

StatementGraphRAG (SGRAG) designates a family of retrieval-augmented generation methodologies that integrate the fine-grained semantics of “statements” within a graph data model to unify, structure, and retrieve knowledge across heterogenous sources for advanced LLM reasoning. The SGRAG paradigm leverages a Statement Graph abstraction, formalized as a directed acyclic graph (DAG) whose vertices and edges encapsulate either RDF-style triples or labeled property graph constructs, and is distinguished by its ability to support lossless interoperability, cross-model querying, and stack-independent data exchange between RDF, LPG, and hybrid graph models (Gelling et al., 2023). SGRAG implementations have been demonstrated both as standalone data management solutions and as subcomponents in state-of-the-art retrieval pipelines for multi-hop, cross-domain, and context-sensitive question answering.

1. Formal Foundations: Statement Graphs as Unifying Abstraction

SGRAG is grounded in the Statement Graph model, which formalizes graphs as directed, vertex- and edge-labeled DAGs. Formally, a Statement Graph consists of a tuple

$G = (V, E, \tau)$

where:

• $V$ is a finite set of vertices,
• $$E \subset V \times \{\mathsf{subject},\,\mathsf{predicate},\,\mathsf{object}\} \times V$$ is a finite set of directed, role-labeled edges,
• $\tau: V \to S \cup E_1 \cup \ldots \cup E_k$ is a total injective function labeling each vertex with a statement identifier from $S$ or a concrete domain element (e.g., IRI, literal, label).

Two vertex categories are defined:

• Leaf vertices: Carry concrete values from domain spaces and have no outgoing edges.
• Internal statement vertices: Represent statements and possess exactly three outgoing edges labeled \emph{subject}, \emph{predicate}, and \emph{object}. For $v$ an internal node, $\tau(v)\in S$, the model enforces:
  • $(v,\mathsf{subject}, v_s)$ with $\tau(v_s)\in S\cup(\bigcup_i E_i)$,
  • $(v,\mathsf{predicate}, v_p)$ with $\tau(v_p)\in(\bigcup_i E_i)$,
  • $(v,\mathsf{object}, v_o)$ with $\tau(v_o)\in S\cup(\bigcup_i E_i)$.

This structure allows for recursive nesting of statements and supports both RDF-star style statement reification and LPG property edges in a uniform formalism. The DAG constraint (no cycles) ensures that higher-level statements reference atomic facts or other statements, with explicit provenance and lineage.

2. Interoperability and Bidirectional Mappings

A central goal in SGRAG is bridging the abstraction gap between RDF-triple models and Labeled Property Graphs, each of which comes with distinct feature sets, ecosystem conventions, and query languages.

The system of bidirectional mappings introduced in the Statement Graph model enables conversion between:

• RDF-Image: Encodes classic RDF (and RDF-star) semantics by constraining statement component types (allowing statements-as-subject/object for reification).
• LPG-Image: Encodes property graph structure, distinguishing edges (entity-to-entity), node labels, and property assignments, with dedicated mapping functions such as:
  • $\mathit{literalvalue}: L \to V$,
  • $\mathit{blank2node}: B \to N$,
  • $\mathit{IRI2node}: I \to N$, $\mathit{IRI2prop}: I \to P$, $\mathit{IRI2label}: I \to 2^{\mathcal{K}}$.

Bidirectional conversion is lossless within the scope of each model’s expressivity, supporting full round-trip exchanges. This enables cross-model graph management: for example, ingesting RDF (e.g., N-Quads/Turtle), mapping to the statement graph, then exporting to LPG (e.g., GraphSON/Cypher) or vice versa.

3. Query Semantics and Cross-Model Reasoning

The unified Statement Graph unlocks cross-model querying capabilities. Because the graph instance can be interpreted as either an RDF-Image or LPG-Image, queries may be posed in any supported language:

• SPARQL/SPARQL-star: For RDF/semantic web workloads.
• Gremlin/openCypher: For property graphs and engineering-centric scenarios.

This is realized by dynamically applying the corresponding image mapping: the underlying Statement Graph is transformed on-the-fly and queries are executed in the expected semantics of the surface model. The semantics for read queries are implicitly defined by the mapping functions, providing model-agnostic access.

Deployment scenarios include multi-tenant knowledge platforms, cross-stack integration services, and systems requiring flexible query language frontends.

4. System Implementation: The 1G Playground

A practical instantiation of SGRAG principles is the “1G Playground” (Gelling et al., 2023), an in-memory DBMS embodying the OneGraph paradigm. The system is designed as follows:

• The central store is the Statement Graph (OneGraph-Image), capable of ingesting both RDF and LPG formats.
• The system exposes two derivative projections:
  • An RDF store (e.g., backed by RDF4J).
  • An LPG store (using Apache TinkerPop).
• The Playground provides interactive interfaces (REPL, REST API) and a shared serialization format (“1G” syntax), supporting both stack-dependent and stack-independent representations.

A single data load yields an integrated graph, which may be queried via SPARQL or Gremlin. The system demonstrates the feasibility of real-world, cross-model, and toolchain-independent graph analytics—supporting import, export, and cross-querying on heterogeneous graph data.

5. Stack-Independent Data Exchange and Serialization

SGRAG’s model enables true graph stack independence: exchange and transformation of graph datasets without committing to a single technology or serialization. Core features include:

• Unified Serialization (“1G” syntax): Encodes statements as a collection of DAG nodes and role-labeled edges with fixed subject/predicate/object semantics, suitable for export/import scenarios.
• Format Flexibility: Ingestion from Turtle, N-Quads, GraphSON, and others is mapped to the canonical statement structure, decoupling data representation from native stack constraints.
• Interoperability Layer: Acts as a mediating schema for pipelines that bridge silos—e.g., integrating RDF-based knowledge graphs with operational LPGs in analytics platforms.

This approach simplifies cross-vendor, cross-standard data mobility, and is foundational in distributed or multi-partner knowledge graph applications.

6. Example Applications and Use Cases

Several applied scenarios motivate and benefit from the SGRAG framework:

• Semantic Data Integration: Combining high-performance LPG-modeled event data with RDF-based reference data in domains such as travel, logistics, or scientific research. For instance, merging flight data (LPG) with DBpedia restaurant facts (RDF) enables advanced geospatial reasoning without semantic compromise.
• Process and Event Analytics: Modeling event transitions (with properties on edges) and their reifications, e.g. in Event Knowledge Graphs where RDF-star reification is harnessed alongside property edge patterns of LPG, supporting analytics on causality and event chains.
• Cross-Domain Data Aggregation: Integrating social, operational, and domain-specific graphs for analytics or machine learning, where each source may natively adopt a different graph standard.
• Tooling and Exchange Ecosystems: Developing external tools, ETL, or conversion utilities that utilize the unified statement model as a neutral data substrate, enabling the development of platform-agnostic graph applications.

The SGRAG principles have been validated in real scenarios via the 1G Playground and other prototype systems.

7. Broader Impact and Future Directions

SGRAG’s abstraction paves the way for a new class of graph management and retrieval-augmented generation systems capable of unifying the strengths and semantic nuances of both RDF and property graphs (Gelling et al., 2023). The mathematical rigor (injective vertex labeling, enforced edge roles, theory-grounded mappings), cross-query interoperability, and practical demonstration in a real system collectively set a foundation for multi-paradigm, multi-lingual, and multi-domain graph analytics.

Emerging research is likely to extend SGRAG toward hybrid reasoning over both richly structured semantic graphs and unstructured data, increase adoption in industrial and scientific integration workflows, and drive adoption of stack-independent exchange practices for federated and explainable knowledge systems.