Home / Architecture

Reference architecture

The ontology control plane for enterprise AI

A vendor-neutral blueprint we adapt to your stack: the ontology becomes the shared meaning layer and the place where AI data policy is defined, enforced and audited.

Layered view Animated flow Decision sequence Deployment

Layered view

Six layers. One source of meaning.

Nothing reaches an AI application without passing through the ontology and the policy decision point. Your source systems stay where they are.

ConsumersAI applications

CopilotsRAG assistantsAutonomous agentsMulti-agent workflowsAnalytics & BIPartner APIs

AccessAI access layer

MCP serversTool gatewayHybrid retrieval APIPrompt assemblySPARQL / GraphQL

EnforcementSemantic policy decision point

Role · attribute · relationshipPurpose & consentMasking & minimisationALLOW / MASK / DENYDecision log

The coreEnterprise ontology + knowledge graph

Domain modules (OWL 2)Sensitivity & purpose (DPV)Policies (ODRL)SHACL shapesResolved entitiesPolicy-filtered vector index

IntegrationSemantic integration

R2RML / RML mappingsLLM-assisted extractionVirtualised queriesCDC & event streamsEntity resolution

SourcesYour systems — unchanged

Operational DBsWarehouse / lakehouseDocuments & emailSaaS appsIoT & logs

CROSS-CUTTINGGovernance & trust

W3C PROV-O lineageImmutable audit logOntologyOps CI/CDCQ & SHACL test suitesEvaluation harnessLeakage red-teamingChange-impact analysisData catalogue syncModel & prompt registry

In motion

Watch facts flow — and get stopped

Blue facts are typed by the ontology; green ones are allowed; amber are masked; red never cross the gate.

Raw dataSemantically typed factAllowed to AIMasked / minimisedDenied at the gate

Decision sequence

What happens in the 40 milliseconds before the LLM answers

Request

User or agent asks a question or calls a tool, with identity, role and declared purpose.

Interpret

The query is mapped to ontology concepts — not just keywords — so intent is explicit.

Plan

Graph traversal, vector and keyword retrieval are planned over the ontology.

Decide

Each candidate fact is checked: sensitivity × purpose × relationship × consent.

Shape

Allowed facts pass; others are masked, generalised or dropped before the LLM sees them.

Generate

The model answers from permitted, typed facts with their evidence paths attached.

Record

Decision, policy version and provenance are logged for audit and evaluation.

Deployment

Runs where your data is allowed to live

Your cloud

Deployed in your AWS, Azure or GCP account and VPC. Managed graph services such as Amazon Neptune or self-hosted Neo4j / GraphDB.

On-premises / air-gapped

Containerised components with no external calls; local LLMs where required for classified or regulated workloads.

Hybrid

Ontology and policies centrally governed; enforcement points run close to each data domain or business unit.

Graph stores

Neo4j · GraphDB · Stardog · Neptune · Fuseki

Vector & search

pgvector · OpenSearch · Pinecone · Weaviate

Agent frameworks

LangGraph · CrewAI · LlamaIndex · MCP

Policy engines

SHACL · OPA · Cedar · custom PDP