# Knowledge graph and ontology for agentic AI: how DataNature gives multi-agent systems semantic context

> An ontology defines the entities your business runs on (customers, accounts, devices, policies, suppliers) and how they relate. DataNature builds a governed knowledge graph from that ontology over Gold tables and documents, so agents can follow relationships instead of guessing joins. Voir agents use it for multi-step work such as deep research, investigations and relationship briefings, with every hop governed and cited.

Source: https://datanature.ai/blog/knowledge-graph-ontology-for-agentic-ai
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

Agents fail on enterprise questions for a simple reason: the answer is spread across entities that are connected in ways no single table shows. Which suppliers share a director with a sanctioned company? Which customers use the same device as an account we closed for fraud? Which policies does this contract clause affect? DataNature answers these with an ontology and a knowledge graph built on governed lakehouse data, and Voir agents use them as context.

## What is an ontology, and how is it different from a semantic layer?

A **semantic layer** defines metrics: what "revenue" or "active customer" means and how to compute it. An **ontology** defines meaning and relationships: what a customer, account, device, merchant or contract is, which attributes identify it, and how it relates to other entities (owns, uses, pays, guarantees, supersedes). The two work together. Cubes answer "how much", the ontology answers "what is connected to what", and documents answer "what does it say".

## What is the knowledge graph built from?

| Source | Becomes |
|---|---|
| Gold tables (customers, accounts, transactions, policies, suppliers) | Entities and relationships, keyed by resolved identities from Silver |
| StreamHouse events (logins, transfers, device registrations) | New edges in near real time, such as "device used by account" |
| DocuContext extractions (contracts, KYC files, claims, procedures) | Entities and relationships found in documents, linked back to the source page |
| Semantic-layer cubes | Metrics attached to entities, such as balance, exposure or claim totals |
| Atlas catalog and lineage | Where each fact came from and who may see it |

Because the graph is built from governed data, it inherits the same controls: masks, row filters and document partitions apply to what an agent can traverse.

## How do agents use the graph?

Graph context changes what an agent can do in three ways:

1. **Grounded traversal.** Instead of writing a multi-join query from scratch, the agent asks the graph for entities within two hops of a customer, then pulls metrics for each from cubes.
2. **Disambiguation.** "Acme" might be three suppliers; the ontology resolves which one the user means from context and identifiers.
3. **Explainable answers.** Every hop is a governed relationship with a source, so the answer can show the path: customer → shared device → closed account → fraud case.

This is often called GraphRAG: retrieval that follows relationships, not only text similarity.

## What does a deep-research multi-agent workflow look like?

Deep research is where multi-agent design pays off. In Voir, a typical pattern uses specialised twins coordinated by a lead agent:

| Agent | Responsibility | Tools it may use |
|---|---|---|
| Planner | Breaks the question into sub-questions and a plan | None beyond the plan |
| Graph researcher | Finds relevant entities and relationships | Knowledge graph, scoped to its role |
| Data analyst | Computes figures for the entities found | Semantic-layer cubes over MCP |
| Document researcher | Finds supporting passages | DocuContext partitions in Milvus |
| Verifier | Checks every claim against a source and flags gaps | Read access to the cited sources |
| Writer | Drafts the report with citations | The verified findings only |

The Agent Meta-Harness enforces each agent's scope inside the tool call, so the document researcher cannot query tables and the analyst cannot read restricted partitions. Actions that change systems, such as opening a case or sending a report, wait for human approval.

## Which use cases benefit most?

- **Fraud and AML investigations**: follow money, devices and beneficiaries across accounts to find rings and mule networks. See [graph analytics for fraud rings and AML](/blog/graph-analytics-fraud-rings-aml-networks).
- **KYC and KYB due diligence**: build ownership and control structures from registries, KYC files and transactions.
- **Relationship briefings**: show a banker everything connected to a client, including guarantors, related companies and open cases.
- **Supplier risk**: map supplier dependencies, certificates and quality incidents across plants.
- **Regulatory impact analysis**: trace a regulation through policies, procedures, controls and the systems that implement them.
- **Network operations**: relate alarms to sites, equipment and customers affected.
- **Deep research reports**: competitive, credit or market research that combines internal data and documents with citations.

## How is the graph governed?

- **Same identities and policies** as the lakehouse: masked attributes stay masked in graph answers.
- **Scoped traversal**: each Voir twin can only traverse entity types and partitions in its scope.
- **Provenance**: every node and edge carries its source table, event or document page.
- **Audit**: graph queries made by agents are logged with the answering twin and user.
- **In-jurisdiction**: the graph, embeddings and agent logs run inside the same deployment as the data.

## Frequently asked questions

### Do we need to design a full enterprise ontology first?

No. Start with the entities one use case needs, such as customer, account, device and transaction for fraud, and extend the ontology as new use cases arrive. The ontology lives alongside the semantic models, so it grows with the platform.

### Is the knowledge graph a separate copy of the data?

It is built from governed Gold tables, events and document extractions, keeps links back to them, and refreshes as they change. The lakehouse remains the system of record.

### How is GraphRAG different from vector RAG?

Vector RAG finds passages similar to the question. GraphRAG follows explicit relationships between entities. Enterprise questions usually need both: relationships to find the right entities, and passages and metrics to answer about them.

## Sources

- [DataNature documentation: Agent Meta-Harness](https://docs.datanature.ai/docs/developers/ai-and-analytics/agent-meta-harness)
- [DataNature documentation: semantic layer](https://docs.datanature.ai/docs/developers/data-management/data-warehouse/semantic-layer)
- [DataNature: a semantic layer for AI agents](/blog/semantic-layer-for-ai-agents-rest-mcp)
- [DataNature: private RAG on governed lakehouse data](/blog/private-rag-on-governed-lakehouse-data)

## How does DataNature Atlas keep this secure and compliant?

Agents are governed like users: scoped, masked, logged and unable to act without approval. [DataNature Atlas](/platform/atlas) is the platform's governance module:

- **Knows your data**: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
- **Controls access**: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
- **Governs AI**: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
- **Proves it**: lineage from source to metric to decision, and an audit log of who or what accessed which data.
- **Stays local**: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.

These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read [DataNature Atlas in depth](/blog/datanature-atlas-governance-rbac-abac-pii-masking) or see [security and compliance](/security).
