Skip to main content

Govern knowledge graphs with schemas

Knowledge

A graph schema is the vocabulary a knowledge source's domain graph promises to keep: which node types exist, which properties identify each node, and which relationships may connect them. You author schemas on the Graph schemas tab of Knowledge Management and attach a published version to a source, so that every load into that source's domain graph is checked against it.

Who can use it: the Admin role or the Knowledge Admin grant, in a tenant licensed for knowledge, in Control Room under Knowledge. The same roles open a source's Domain graph tab.

How it works​

A source's domain graph​

Every source's detail page has a Domain graph tab: a separate, typed layer of nodes and relationships that agents query with a graph query tool. It does not change search ranking, and the source's documents are searchable regardless. The tab has four views:

  • Ontology, the default: one card per node type with its count, properties and sample values, then a Relationships list with edge counts.
  • Canvas: an interactive drawing of the graph with Search the graph…, a Node types legend, Fit, Re-layout, saved Views and Export. Select a node to inspect it, double-click to expand it.
  • Explore: pick a Node type, add property filters and select Run query; results are capped at 50 rows, and a row shows its Connections one hop away.
  • Schema: the source's schema posture, described below.

A graph fills in one of two ways. When a schema is attached, the platform extracts typed nodes and relationships from the source's documents during a reindex, which attaching a schema starts for you. Alternatively, Load data takes JSON files of nodes and edges from your own extraction pipeline, with Replace the existing graph choosing between a refresh and a merge. Until then the tab says No graph data and offers Set up a schema.

What a graph schema is​

A schema declares Node types, each with a label, an optional Display name and properties (a name, a type of string, number, boolean or date, and whether it is part of the Key, is Required or has Allowed values), and Relationships, each a relationship type from one node type to another. Validation is open-world: properties on loaded data that the schema does not declare are always fine.

A schema is a series of numbered versions. A Draft is editable; a Published version is immutable and sources attach that exact version. To change a published schema you create a New version, which clones it into a fresh draft. The Graph schemas tab lists one row per schema with its Versions, Node types, Relationships and Updated time, filtered by All, Drafts or Published.

How a schema is attached and checked​

The Schema view of a source's Domain graph has three cards. Attached graph schemas lists the attached versions; pick one under Pick a published schema… and select Attach, or detach one from its chip. Publishing a newer version changes nothing until you attach it. Enforcement is Off (loads are accepted as they are), Advisory (loads are validated, always accepted, and violations are reported: the safe default) or Strict (a load with any violation is rejected whole). Conformance shows the last evaluated load: nothing yet, Last load conforms to the attached versions, or a count of violations with a report grouped by rule (such as Unknown node label, Missing key property or Unknown relationship) with examples. Acknowledge & baseline clears the badge while keeping the report.

The same posture appears as a pill in the Domain graph header and on the Sources table for rows that carry a domain graph.

The knowledge engine's Entity graph setting is a different thing: entity extraction at ingest that boosts or injects related passages at query time, separate from any source's domain graph. See Tune the knowledge engine.

Create, publish and attach a schema​

  1. Start a draft

    On the Graph schemas tab, New schema opens an empty editor. New from source opens Promote a schema from a source: choose a Source with a domain graph and select Discover schema. The source's loaded graph (labels, properties, observed values) becomes an editable draft, and nothing is saved until you save it. A source's Ontology view offers the same as Promote to graph schema.

    The Graph schemas tab of Knowledge Management: the filters All, Drafts and Published, the buttons New from source and New schema, the empty state Govern your knowledge graphs with a schema with a New schema button, and the Next steps links Attach a schema on a source's Schema tab and Explore a loaded graph.The Graph schemas tab of Knowledge Management: the filters All, Drafts and Published, the buttons New from source and New schema, the empty state Govern your knowledge graphs with a schema with a New schema button, and the Next steps links Attach a schema on a source's Schema tab and Explore a loaded graph.
    The Graph schemas tab before any schema exists. Filter by All, Drafts or Published; New schema starts an empty draft, and New from source promotes a draft from a source that already carries a domain graph.

    A source's Schema view adds a Create a schema card: Generate a schema from your documents drafts one from the source's text, and Import schema (JSON) takes your own file; both publish and attach the schema in one action with Advisory enforcement and start a reindex.

  2. Define node types and relationships

    Give the draft a Schema name and a Description. Add node type adds a card: type its label, add properties, tick Key on those that identify a node and Required on those every node must carry, and list Allowed values for coded fields. Add relationship takes a from node type, a relationship type and a to node type. The Fix before saving: banner lists what still fails; Save draft stores the draft and Cancel leaves the editor.

    The schema editor for a new draft: the breadcrumb Graph schemas / New schema, the Schema name field with a Draft badge, the buttons Cancel, Save draft, Publish and Test a payload, the Fix before saving list, the About card with Description, and Node types (1) with Add node type and one node type card showing Label, Display name, a property row with property name, type, Key, Required and Allowed values, and Add property.The schema editor for a new draft: the breadcrumb Graph schemas / New schema, the Schema name field with a Draft badge, the buttons Cancel, Save draft, Publish and Test a payload, the Fix before saving list, the About card with Description, and Node types (1) with Add node type and one node type card showing Label, Display name, a property row with property name, type, Key, Required and Allowed values, and Add property.
    A new draft in the schema editor. Fix before saving lists what the definition still needs; each node type card takes a label, its properties and which of them form the key. Save draft keeps it editable, Publish freezes the version.

    Test a payload validates a pasted payload against the definition without writing anything.

  3. Publish

    Select Publish and confirm. The confirmation reminds you to attach it to a source to enforce it.

  4. Attach and choose enforcement

    Open the source, then Domain graph, then Schema. Pick the version, select Attach and set Enforcement. Start with Advisory, and move to Strict once Conformance reports a clean load.

Limits and behaviour​

  • Labels, property names and relationship types start with a letter or underscore, use only letters, digits and underscores, and are at most 100 characters. entity, chunk and document are reserved labels; relates_to, mentions and mentioned_in are reserved relationship types.
  • Every node type needs at least one property and at least one key property; Allowed values apply to string properties only; a relationship must point at declared node types.
  • Deleting a draft removes it permanently; deleting a published version archives it, and a version attached to a source cannot be deleted until it is detached. A schema has one open draft at a time.
  • Attaching a schema whose declarations clash with one already attached is refused. Strict has no effect while nothing is attached.
  • A payload file holds at most 10,000 nodes and 20,000 edges, and one load takes up to 50 files; a load that stops part way resumes from the failed file with Retry from.
  • When the graph backend is unreachable the tab says Graph unavailable; search on the source is unaffected.

Describes the platform as of .

Was this page helpful?