What is contextualization?
Contextualization is the work of attaching each piece of data to the real-world thing it describes: this signal belongs to that pump, this work order concerns that vessel, this datasheet documents that transformer.
It sounds mundane, and it is the single activity that turns a pile of records into a model you can reason over. The ontology describes what can exist; the knowledge graph is the result; contextualization is the doing.
The problem, in one picture
The same physical pump is known to four systems under four different names, and none of them records that the four are the same thing:
Everything DataHub does downstream, traversal, investigation, reporting, agents, the digital twin itself, assumes those links exist. Contextualization is how they come to exist. It is the bridge between data liberation, which makes the records reachable, and the knowledge graph, which makes them navigable.
How it relates to the other three words
| Word | What it is | Analogy |
|---|---|---|
| Taxonomy | The classes things can belong to | The library's subject scheme |
| Ontology | Classes plus the relationships allowed between them | The cataloguing rules |
| Contextualization | The activity of linking real records to real things | The librarian cataloguing each book |
| Knowledge graph | The populated result | The finished catalogue |
| Digital twin | The graph, kept in step with live data | The catalogue, plus live circulation records |
The first two are decided by a few people in a room and revisited rarely. Contextualization is continuous: every new signal, document and source system arrives uncontextualized, and the value of the platform tracks how quickly it gets connected.
What a business actually gets from it
Each of these is impossible while the links live in people's heads, and routine once they are data:
- Search that finds things. Ask for a pump and get its signals, events, documents and neighbours, not four separate result lists in four systems.
- Cross-silo answers. "Which assets had maintenance open during the pressure excursion?" is only answerable when work orders and signals point at the same resources.
- Faster onboarding. A new engineer navigates the model instead of spending a year learning which tag maps to which equipment number. The mapping stops being tribal knowledge, which is also what removes the key-person risk.
- Documents attached to equipment. The datasheet is reachable from the pump, rather than from whoever remembers the folder. Files →
- Usable AI. An agent's accuracy is set by the context available to it. Contextualization is, quite literally, the manufacturing of that context. And once several agents run as an organisation, it is what lets all of them mean the same thing by "pump P-101".
How it is done in practice
Mostly by exploiting identifiers your operation already maintains, which is why naming discipline matters so much.
A plant tag such as COM-99-PT-1034 already encodes facility, system and instrument.
When the historian, the maintenance system and the model all carry the tag somewhere,
linking them is mechanical. This is the workhorse method, and the reason to mirror your
tagging standard in the external id.
Mirroring it is literal: DataHub stores the tag exactly as you send it, uppercase, hyphens and all, so every system holds the same string and the match is a comparison rather than a normalisation each integration has to reimplement the same way.
When systems disagree, the tag prefix, the location code or the system number usually narrows a record to a handful of candidates, and a person confirms the match once. The confirmed link is then permanent.
Some fraction always needs judgement, a renamed unit, a retagged loop, a document with no reference on it. This is domain-expert work in the console, and it is the part worth scheduling rather than hoping away.
Once the model exists, every new source is linked as it lands, so the backlog never rebuilds. An integration that writes data without relating it to resources is leaving the job half done. CDC and landing data well →
For a facility with a consistent tagging standard, the bulk of contextualization is automatable and the manual remainder is small. For an operation whose systems grew without shared identifiers, the manual share is larger, and that effort is the project. Assess which situation you are in before promising a timeline, it is the single best predictor of integration cost.
On the roadmap, this is also where assisted classification lands: suggesting the ontological placement of new or unfamiliar data, so the human confirms instead of constructs. AI agents and integration work →
- Naming and standards: the identifiers matching depends on
- The knowledge graph: what contextualized data makes possible
- Data liberation: the step before this one
- Building your model: where the classes and relationships come from
- The cost of doing nothing: what uncontextualized data costs