Open Data Infrastructure
The Context Graph: Metadata, Lineage, Semantics, and Policy for AI
Introduce the context graph as connected metadata: lineage, semantics, policy, and provenance that make AI outputs defensible.
When an AI system answers a question, you do not just want an answer. You want the why behind the answer. That is what a context graph gives you.
Why it matters
Without a context graph, AI systems fail in quiet ways. They pick the wrong table. They use stale data. They ignore policy because it is not encoded in a way the system can evaluate.
A context graph is how you teach a system what things mean, how they relate, and which constraints apply. It is connected metadata with edges that matter.
The ODI angle
ODI makes a context graph possible because it pushes metadata, lineage, and policy into open interfaces.
If your context graph depends on a closed vendor metadata model, it will be hard to reuse across agents and tools. You will end up building five versions of the same thing.
The point is not another graph database. The point is a shared, governed representation of meaning that tools and agents can query.
Core idea: a context graph is the bridge between data and defensible AI outputs.
The architecture test
For AI architects, the test is whether the graph helps humans and machines reason the same way.
- Define core entities: tables, columns, metrics, owners, policies, and lineage edges.
- Use open lineage standards where possible.
- Treat semantic models as contracts, not documentation.
- Store policy in a form tools and agents can evaluate.
- Make the graph queryable for incident response and audits.
What breaks first
This breaks when metadata lives in disconnected tools with inconsistent truth.
- Metadata lives in five disconnected tools, so the AI system gets conflicting signals.
- Lineage exists after the incident, not during design and review.
- Semantics live in docs retrieval cannot interpret.
- Policy is enforced in apps but not in data and tool paths.
Questions to ask
Use these questions when you evaluate context graph AI as a real infrastructure layer.
- Which sources feed your context graph, and how do you reconcile conflicts?
- How do you keep lineage current as pipelines change?
- Where do semantic definitions live, and how do you version them?
- How do you evaluate policy consistently across tools?
- Can an agent query the graph and explain a result with citations?
If the graph is only a UI feature, it will not help your agents. Agents need a queryable contract.
Sources to start with
Start with open metadata and lineage systems, then define the contract you want to preserve.