Fabric Graph

Microsoft Fabric Graph (GraphModel item)

Graph in Fabric models a labeled property graph directly over OneLake Delta tables — no ETL, no data duplication. You define node/edge types and map them to columns; on save, Fabric ingests the tables and builds a read-optimized queryable graph. Query it with GQL (the ISO/IEC 39075 standard) via UI, REST, or NL2GQL.

Item type name: GraphModel (this is the fabric-cli .GraphModel suffix and the REST collection /GraphModels). New item → Analyze and train data → Graph model.

Three things that catch people out

1. It is NOT the KQL graph operators. Fabric Graph (GraphModel) is a separate workload from the make-graph / graph-match / #crp query_language=gql story in Eventhouse/KQL (see fabric-eventhouse). Same word "graph", different engine, different API, different syntax surface. Don't cross-apply.
2. GQL is read-only here. You cannot INSERT / SET / DELETE graph data through GQL. Data is loaded and refreshed via data management (save the model = reingest from OneLake). CREATE GRAPH is only partially supported and DROP GRAPH is not in GQL — drop via the Fabric UI or REST API instead.
3. No schema evolution. Once nodes/edges/properties are modeled and data is ingested, the structure is fixed. Adding a property, changing a label, or changing a relationship type means reingesting source data into a new model. Plan the schema up front.

GQL query language (ISO/IEC 39075)

GQL expresses relationships as visual patterns instead of joins. Roots are SQL + openCypher, but it is its own dialect — do not assume Cypher syntax (e.g. no CREATE (n:Label) data writes here).

MATCH (c:Customer)-[:purchases]->(o:`Order`)
RETURN c.fullName AS customer_name, count(o) AS num_orders
GROUP BY customer_name
ORDER BY num_orders DESC
LIMIT 5

• Backtick-quote labels that collide with keywords: ` (o:Order) `.
• || is string concatenation: person.firstName || ' ' || person.lastName.
• Undirected match -[:knows]- is allowed in patterns, but the graph itself has no undirected edges — every edge is directed and has exactly one label.
• Comments: -- line, // line, /* */ block.

Graph-type DDL (defining the schema in GQL)

The graph type (schema) uses a compact DDL that is unlike anything in training-data Cypher. Get this exact, or a model will invent invalid syntax:

-- node type with typed properties; :: is the type operator
(:Person => {
    id :: UINT64 NOT NULL,
    firstName :: STRING,
    birthday :: UINT64
}),

-- key constraint (node identity)
CONSTRAINT person_pk FOR (n:Person) REQUIRE (n.id) IS KEY,

-- inheritance with =>  (abstract base, then subtypes)
ABSTRACT (:Place => { id :: UINT64 NOT NULL, name :: STRING }),
(:City => :Place),
(:Country => :Place),

-- property extension with +=
(:Post => :Message += { language :: STRING, imageFile :: STRING }),

-- edge types: source -[:label {props}]-> destination
(:Person)-[:knows { creationDate :: ZONED DATETIME }]->(:Person),
(:Person)-[:studyAt { classYear :: UINT64 }]->(:University),

-- ( 19800101 RETURN n.firstName, n.lastName LIMIT 100" }'

• Token audience is https://api.fabric.microsoft.com (az account get-access-token --resource https://api.fabric.microsoft.com). See fabric-auth.
• List graphs in a workspace: GET /v1/workspaces/{workspaceId}/GraphModels.
• Application errors return HTTP 200, not 4xx. You MUST inspect the GQL status.code (6-char string): 00/01/02/03 prefixes = success (possibly with warnings); anything else is an error, with detail in status.diagnostics and chained status.cause. Do not assume success from HTTP 200.
• Result is a discriminated union on result.kind (TABLE with columns+data, or NOTHING). Values carry gqlType; large INT64/UINT64 outside JS safe range and float Inf/-Inf/NaN arrive as strings.

Request/response schema, value-encoding table, and status-code families → [references/REFERENCE.md](references/REFERENCE.md).

Graph Model item-definition JSON (automation / CI-CD)

For createItemWithDefinition / getDefinition / updateDefinition (see fabric-tmdl-api for the envelope, LRO, base64, all-parts rules), a GraphModel definition has four required parts:

| Part | What it holds | |---|---| | dataSources | DeltaTable sources: { name, type:"DeltaTable", properties:{ path: "abfss://…/Tables/" } } | | graphDefinition | nodeTables + edgeTables mapping source columns → graph: propertyMappings, and on edges sourceNodeKeyColumns / destinationNodeKeyColumns; optional filter (Single/Group with and/or, operators Equal/Contains/GreaterThan/In…) | | graphType | nodeTypes (alias, labels[], primaryKeyProperties[], properties[]) + edgeTypes (alias, labels[], sourceNodeType.alias, destinationNodeType.alias, properties[]) | | stylingConfiguration | modelLayout: node positions, styles (size), pan, zoomLevel — visual only |

Note the indirection: the graphType declares abstract aliases (Customer_nodeType) and labels; the graphDefinition binds each alias to a dataSourceName and maps sourceColumn → propertyName. A composite key is expressed as multiple primaryKeyProperties and matching destinationNodeKeyColumns. Full JSON examples for all four parts → [references/REFERENCE.md](references/REFERENCE.md).

Data load & refresh

• Save = ingest. Selecting Save in the model editor both persists the model and reingests OneLake data to rebuild the queryable graph. To pull fresh data without model changes, just save again.
• Scheduled refresh: workspace → graph item → … → Schedule → Add schedule (minute/hourly/daily/weekly/monthly). Not available in My Workspace — shared workspaces only.

Other gotchas / limits

• Storage floor 100 GB provisioned (billed at OneLake Cache rate); compute billed by CPU uptime at 10 CU-seconds per second, rounded up to the minute. Uses your existing Fabric capacity — no separate SKU.
• Hard limits: max 10 graph models per workspace; the GA SKU processes roughly 2 billion graph elements (nodes + edges — contact the product team for larger); variable-length patterns support up to 8 hops; queries time out at 20 minutes and responses are truncated above 64 MB (aggregation gets unstable past 128 MB). String property max 65,535 chars; List property max 65,535 elements. See limitations.
• NL2GQL (natural-language → GQL via Fabric Data Agent, see fabric-data-agent) is preview. openCypher support is preview and is the KQL-graph path, not this item.
• Set operations not yet supported. UNION DISTINCT, EXCEPT, INTERSECT, and OTHERWISE are not available — compose with linear chaining of core statements (MATCH/LET/FILTER/RETURN) instead. Full conformance-gap list → limitations.
• Governed by OneLake security + workspace RBAC (see fabric-security).

Reference

• MS Learn: Graph overview · How graph works · GQL language guide · GQL schema example · GQL conformance
• MS Learn: GQL Query API reference · Graph Model definition (REST) · Manage and refresh data
• Consolidated link bundle, full schema/JSON/value-encoding tables: [references/REFERENCE.md](references/REFERENCE.md)

See also

• fabric-eventhouse — the other graph story: KQL make-graph / graph-match / openCypher in Eventhouse. Different engine; do not conflate.
• fabric-tmdl-api — definition envelope, LRO polling, base64, "include all parts" rule (applies to GraphModel definitions too)
• fabric-rest-api — runtime item ID vs .platform logicalId, LRO 202 pattern, pagination
• fabric-auth — api.fabric.microsoft.com token audience for the query API
• fabric-data-agent — graph as a Data Agent source for NL2GQL (preview)
• fabric-cli — fab path syntax Workspace.Workspace/Item.GraphModel, export/import

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: wardawgmalvicious
• Source: wardawgmalvicious/claude-config
• License: MIT

Install and usage instructions live in the source repository linked above.