Graph-Based Verification

This feature validates that generated statements conform to the graph schema (ontology). It checks for impossible relationships, such as connecting entities of incompatible types (e.g., a Person cannot manufacture a ChemicalElement). This prevents category errors and ensures outputs respect predefined domain and range constraints.

• Example: A graph schema defines that the employedBy relationship only exists between a Person and a Company. The system would flag the statement "The Eiffel Tower is employed by Google" as structurally invalid.

Verification assesses plausibility by checking for the existence of connecting paths between entities mentioned in a generated claim. A claim is more plausible if a short, meaningful path exists in the knowledge graph.

• Direct Edge Check: Verifies if a specific relationship (predicate) exists directly between two entities.
• Multi-Hop Path Discovery: For complex claims, the system searches for indirect connections. The claim "Molecule X treats Disease Y" is supported if paths exist like (X) --[targets]--> (Protein Z) --[involvedIn]--> (Y).
• Path Absence as a Red Flag: The complete lack of any connecting path strongly suggests the claim is a hallucination or requires external, uncited knowledge.

Knowledge graphs encoded with description logics (via OWL) enable automated logical reasoning. Verification uses these rules to detect contradictions.

• Key Constraint Types:
  • Disjointness: CEO and Intern are disjoint classes; an entity cannot be both.
  • Inverse Relationships: If parentOf is the inverse of childOf, and the graph contains (Alice) --[parentOf]--> (Bob), then the claim "Bob is not a child of Alice" is logically false.
  • Transitivity: If partOf is transitive, and (Engine) --[partOf]--> (Car) and (Car) --[partOf]--> (Fleet), then the claim "The Engine is part of the Fleet" is verified as true.

This transforms the graph into a deductive system for consistency checking.

Verifies that claims align with temporal facts and numerical attributes stored in the graph. This prevents anachronisms and quantitative errors.

• Temporal Verification: Checks if event sequences or entity lifespans are consistent. The claim "Steve Jobs launched the iPhone in 1995" fails against a graph where iPhone has a launchDate of 2007 and SteveJobs has a birthDate of 1955.
• Attribute Bounding: Validates numerical claims against known ranges. If a Product has a price of $999, the claim "It costs over $2000" is flagged. This is crucial for deterministic grounding in financial or scientific domains.

A core feature is source node tracing, which provides an audit trail. Every verified (or flagged) claim can be linked directly to the supporting or contradicting subgraph.

• How it works: The system records the specific nodes, edges, and inferred paths used during verification.
• Output: Generates an explanation such as: "Claim CONFIRMED. Supported by path: (Entity A) --[Relationship R]--> (Entity B), sourced from dataset D."
• Benefit: This creates algorithmic trust by making the verification process transparent, debuggable, and suitable for governance workflows. It answers the "why" behind a verification result.

Graph-based verification is not just a final checkpoint; it actively improves the Retrieval-Augmented Generation (RAG) pipeline through feedback.

• Retrieval Re-ranking: Verification scores can re-rank retrieved subgraphs, prioritizing those that lead to more verifiable outputs.
• Prompt Correction: If a claim fails verification, the system can inject corrective context (e.g., "The graph shows the date was 2007, not 1995") and trigger a regeneration.
• Knowledge Graph Completion: Persistent verification failures on a specific type of claim can signal a gap in the knowledge graph, triggering an update pipeline. This closes the loop between generation and knowledge graph quality assessment.

A factual consistency check is a post-generation verification step that compares a language model's output against a source knowledge graph to identify contradictions.

• Process: The generated text is parsed to extract claimed subject-predicate-object triples. These are then queried against the knowledge graph.
• Outcome: Produces a binary or confidence-scored result indicating if the claim is supported, contradicted, or not found in the verified graph.
• Contrast with Graph-Based Verification: While a consistency check is a specific evaluation task, graph-based verification is the broader methodology using the graph's inherent structure to perform it.

Deterministic grounding is the principle of explicitly linking every generated statement to a verifiable source fact or subgraph within a knowledge graph.

• Core Tenet: It mandates traceability. Every claim must have a provenance trail back to specific graph nodes and edges.
• Implementation: Achieved through techniques like source node tracing and triple attribution in the prompt or output metadata.
• Purpose: Moves beyond probabilistic confidence, providing an auditable, fact-based anchor for generated content. It is the foundational goal that graph-based verification operationalizes.

Knowledge graph quality assessment is the evaluation of a graph's accuracy, completeness, and consistency—the very attributes that determine its efficacy for verification.

• Key Metrics: Includes accuracy (are the facts true?), completeness (are relevant facts missing?), consistency (are there logical contradictions?), and freshness.
• Direct Impact: A verification system is only as reliable as the graph it queries. Schema conformance (validity against an ontology) and link correctness are critical pre-conditions.
• Methods: Involves rule-based validation, crowdsourcing, and triple confidence scoring to establish a 'verifiability baseline' for the graph itself.

Neuro-symbolic RAG is an architecture that integrates neural language models with symbolic reasoning over a knowledge graph, creating a robust framework for verification.

• Components: Combines a neural retriever/generator (for understanding natural language) with a symbolic reasoner (for applying logical rules).
• Verification Synergy: The neural component proposes or generates content; the symbolic component, operating on the graph, validates it. This hybrid approach overcomes the limitations of purely statistical methods.
• Example: A language model suggests a drug side effect. A neuro-symbolic system would check this against a biomedical knowledge graph's formal ontology for drug-contraindication relationships, providing a logically sound verification step.

Source node tracing is an explainability technique that records and presents the specific nodes and edges used to verify or generate a statement.

• Mechanism: During retrieval and verification, the system logs the unique identifiers (URIs) of all graph elements consulted.
• Output: Provides an audit trail, often displayed as a list of source triples or a visual subgraph, showing the exact factual basis for an output.
• Critical for Verification: It transforms the abstract concept of 'graph-based verification' into a tangible, inspectable artifact, allowing engineers and users to confirm the verification process themselves.