azure-kusto-graph

Kusto Graph Semantics

Build transient and persistent graphs from tabular data using KQL graph operators. This skill translates natural language into the edges-first graph construction pattern and graph query operators.

Activation Triggers

Use this skill when the user:

• Wants to build a graph from tabular data (make-graph)
• Asks to find patterns, paths, or relationships in data
• Mentions graph-match, graph-shortest-paths, graph-to-table, graph-mark-components
• Wants to create a persistent graph model or snapshot
• Says "build a graph", "find the shortest path", "find connected components", "show relationships"
• Asks about transient vs persistent graphs

Not a natural-language-to-KQL converter. The input should generally be a working KQL query whose results the user wants converted to a graph, plus a natural-language description of the desired graph structure. Basic NL source requests are supported only when they map directly to a known table with obvious columns. For general NL-to-KQL conversion, use a dedicated query-generation skill (available separately).

Complementary skills:

• azure-kusto-irql -- composable security query primitives that produce the tabular inputs for graphs
• azure-kusto-irql-graph -- IRQL's Lift_To_Graph JSON mapping system for richly-typed, icon-decorated graphs in Kusto Explorer

The Edges-First Approach

The fundamental pattern for building graphs in Kusto:

1. Define your EDGES       -> src --> dest, with relationship type/properties
2. Define your NODE LOOKUPS -> display names, types, properties for each node ID
3. Union edge types         -> if you have multiple relationship types
4. Union node lookups       -> if you have multiple node types
5. Call make-graph          -> edges | make-graph Source --> Target with nodes on nodeId

This is how to think in make-graph. Edges are the relationships you care about. Nodes are lookup tables that give those IDs a face -- display names, types, properties.

Graph Operators Reference

make-graph -- Build a graph from tables

Edges | make-graph SourceId --> TargetId with Nodes on NodeId

• Edges: tabular source where each row is an edge
• SourceId --> TargetId: columns containing source and target node IDs
• with Nodes on NodeId: optional node property table joined by ID
• Supports multiple node tables: with Nodes1 on Id1, Nodes2 on Id2
• Nodes appearing in edges but missing from the node table get empty properties

graph-match -- Find patterns

G | graph-match (a)-[e]->(b) where <constraints> project <output>

Pattern notation:

Multi-hop patterns: (a)-[e1]->(b)-[e2]->(c) Star patterns: (a)--(center)--(b), (c)--(center)--(d) Cycles control: cycles = all | none | unique_edges (default: unique_edges)

graph-shortest-paths -- Find shortest paths

G | graph-shortest-paths (start)-[e*1..20]->(end)
      where start.name == "Alice" and end.name == "Server01"
      project Path = e, Length = array_length(e)

• Requires at least one variable-length edge
• output = any (default, one path per pair) or output = all (all equal-length shortest paths)
• Variable-length edge properties returned as dynamic arrays

graph-to-table -- Export graph to tables

G | graph-to-table nodes                                     // export nodes
G | graph-to-table edges                                     // export edges
G | graph-to-table nodes as N, edges as E                    // export both
G | graph-to-table nodes with_node_id=Id                     // include node hash ID
G | graph-to-table edges with_source_id=Src with_target_id=Tgt  // include edge endpoint IDs

graph-mark-components -- Find connected components

G | graph-mark-components with_component_id=ComponentId
  | graph-to-table nodes
  | summarize Members = make_list(name) by ComponentId

Assigns a ComponentId to each node. Nodes in the same connected component share the same ID.

graph() function -- Query persistent graphs

graph("MyGraphModel")                              // latest snapshot
graph("MyGraphModel", "Snapshot_2025_01")           // specific snapshot
graph("MyGraphModel", true)                         // transient from model definition

Transient Graphs

Created dynamically during query execution. No setup required. Ideal for ad-hoc analysis, exploration, and prototyping.

Template: Basic two-entity graph

// 1. Define edges
let edges = <SourceTable>
    | summarize <aggregations> by SourceCol, TargetCol;
// 2. Define node lookups
let source_nodes = edges
    | distinct SourceCol
    | project nodeId = SourceCol, label = SourceCol, nodeType = "<SourceType>";
let target_nodes = edges
    | distinct TargetCol
    | project nodeId = TargetCol, label = TargetCol, nodeType = "<TargetType>";
let all_nodes = union source_nodes, target_nodes;
// 3. Build and query the graph
edges
| make-graph SourceCol --> TargetCol with all_nodes on nodeId
| graph-match (s)-[e]->(t)
    where <constraints>
    project Source = s.label, Target = t.label, <edge properties>

Template: Multi-relationship graph

// Multiple edge types -> union them with a common schema
let auth_edges = AuthEvents
    | project Source = username, Target = hostname, edgeType = "authenticates", ts = timestamp;
let net_edges = NetworkEvents
    | project Source = src_ip, Target = url, edgeType = "connects", ts = timestamp;
let all_edges = union auth_edges, net_edges;
// Node lookups from all sources
let user_nodes = Employees | project nodeId = username, label = name, nodeType = "User";
let host_nodes = AuthEvents | distinct hostname | project nodeId = hostname, label = hostname, nodeType = "Host";
let all_nodes = union user_nodes, host_nodes;
all_edges
| make-graph Source --> Target with all_nodes on nodeId

Persistent Graphs

For large-scale, reusable graphs. Stored in database metadata. Support snapshots for historical comparison.

Safety: Creating or altering graph models and snapshots modifies the database. Always show the exact command and confirm with the user before executing .create-or-alter graph_model or .make graph_snapshot.

Step 1: Create a graph model

.create-or-alter graph_model SecurityGraph
{
  "Schema": {
    "Nodes": {
      "User": {"name": "string", "role": "string"},
      "Host": {"hostname": "string"},
      "IP":   {"ip": "string"}
    },
    "Edges": {
      "AuthenticatesTo": {"timestamp": "datetime", "result": "string"},
      "ConnectsFrom":    {"timestamp": "datetime"}
    }
  },
  "Definition": {
    "Steps": [
      {
        "Kind": "AddNodes",
        "Query": "Employees | project name, role",
        "NodeIdColumn": "name",
        "Labels": ["User"]
      },
      {
        "Kind": "AddNodes",
        "Query": "AuthenticationEvents | distinct hostname | project hostname",
        "NodeIdColumn": "hostname",
        "Labels": ["Host"]
      },
      {
        "Kind": "AddEdges",
        "Query": "AuthenticationEvents | project username, hostname, timestamp, result",
        "SourceColumn": "username",
        "TargetColumn": "hostname",
        "Labels": ["AuthenticatesTo"]
      }
    ]
  }
}

Step 2: Create a snapshot

.make graph_snapshot SecurityGraph Snapshot_2025_07

Step 3: Query the snapshot

graph("SecurityGraph")
| graph-match (user)-[auth]->(host)
    where user.role == "Admin" and auth.result == "Failed Login"
    project User = user.name, Host = host.hostname, Time = auth.timestamp

Management commands

Safety: All control commands below modify or delete database objects. Never execute .drop, .create-or-alter graph_model, or .make graph_snapshot automatically. Always show the exact command, cluster, database, and affected object, then require explicit user confirmation before execution.

.show graph_models                        // list all models
.show graph_model SecurityGraph           // show model details
.show graph_snapshots SecurityGraph       // list snapshots
.drop graph_snapshot SecurityGraph Snapshot_2025_07  // delete a snapshot (CONFIRM FIRST)
.drop graph_model SecurityGraph           // delete model and all snapshots (CONFIRM FIRST)

Transient vs Persistent: When to Use Which

Security & Threat Hunting Examples

Authentication graph: who logged into what from where

let auth_edges = AuthenticationEvents
    | summarize
        logins = count(),
        fails = countif(result == "Failed Login")
      by src_ip, username, hostname;
let ip_nodes = auth_edges | distinct src_ip
    | project nodeId = src_ip, label = src_ip, nodeType = "IP";
let user_nodes = auth_edges | distinct username
    | project nodeId = username, label = username, nodeType = "User";
let host_nodes = auth_edges | distinct hostname
    | project nodeId = hostname, label = hostname, nodeType = "Host";
let all_nodes = union ip_nodes, user_nodes, host_nodes;
// IP -> User edges
let ip_user = auth_edges
    | project Source = src_ip, Target = username, logins, fails;
// User -> Host edges
let user_host = auth_edges
    | project Source = username, Target = hostname, logins, fails;
union ip_user, user_host
| make-graph Source --> Target with all_nodes on nodeId
| graph-match (ip)-[e1]->(user)-[e2]->(host)
    where e2.fails > 20
    project
        IP = ip.label,
        User = user.label,
        Host = host.label,
        Failures = e2.fails
| order by Failures desc

Lateral movement detection: users sharing compromised hosts

// Pattern: (user1)-[auth1]->(host)<-[auth2]-(user2)
// Two users both failing on the same host = possible credential spray
let edges = AuthenticationEvents
    | summarize fails = countif(result == "Failed Login"), logins = count()
      by username, hostname;
let nodes = union
    (edges | distinct username | project nodeId = username, nodeType = "User"),
    (edges | distinct hostname | project nodeId = hostname, nodeType = "Host");
edges
| make-graph username --> hostname with nodes on nodeId
| graph-match (u1)-[e1]->(h)<-[e2]-(u2)
    where u1.nodeId != u2.nodeId and e1.fails > 10 and e2.fails > 10
    project
        User1 = u1.nodeId, User2 = u2.nodeId,
        SharedHost = h.nodeId,
        User1Fails = e1.fails, User2Fails = e2.fails
| distinct User1, SharedHost, User2, User1Fails, User2Fails
| order by User1Fails + User2Fails desc

Shortest attack path

let edges = SecurityEvents
    | project Source = source_entity, Target = target_entity, action, timestamp;
let nodes = union
    (edges | distinct Source | project nodeId = Source),
    (edges | distinct Target | project nodeId = Target);
edges
| make-graph Source --> Target with nodes on nodeId
| graph-shortest-paths (start)-[e*1..10]->(end)
    where start.nodeId == "ExternalIP_1.2.3.4" and end.nodeId == "DatabaseServer"
    project
        PathLength = array_length(e),
        Actions = e.action,
        Hops = e.Target

Connected components: find isolated clusters

let edges = NetworkFlows
    | project Source = src_ip, Target = dst_ip;
let nodes = union
    (edges | distinct Source | project nodeId = Source),
    (edges | distinct Target | project nodeId = Target);
edges
| make-graph Source --> Target with nodes on nodeId
| graph-mark-components with_component_id = ComponentId
| graph-to-table nodes
| summarize Members = make_list(nodeId), Size = count() by ComponentId
| order by Size desc

Visualize in Kusto Explorer

End a query at make-graph (without piping to graph-match) to trigger Kusto Explorer's interactive graph visualization window:

edges
| make-graph Source --> Target with all_nodes on nodeId
// <- stop here. Kusto Explorer renders the graph visually.

To flatten back to a table for dashboards or export, pipe through graph-match | project or graph-to-table.

Using with IRQL

When working with security data, consider using IRQL selectors (Get_*) from the azure-kusto-irql skill as the data source. IRQL gives you a unified schema without memorizing raw table names or column mappings. For rich visualization with icons and node folding, the azure-kusto-irql-graph skill's Lift_To_Graph is the faster path.

Note: Lift_To_Graph, Graph_Render_View, and Graph_Fold_By_Property are stored functions, not built-in operators. They are pre-deployed on the kc7001 example cluster but may need deployment on other clusters. See azure-kusto-irql-graph/references/DEPLOY_IRQL_FUNCTIONS.md for function definitions and deployment instructions.

Example: IRQL selectors -> make-graph -> shortest path

IRQL handles the data retrieval; make-graph handles the graph analysis. This finds the shortest path from an external IP to a mail server through auth events:

// IRQL provides unified columns (ClientIp, Hostname, Username, Result)
let auth = Get_Event_Authentication_All
    | where Result == "Failed Login";
let edges = auth
    | summarize Failures = count() by ClientIp, Hostname;
let nodes = union
    (edges | distinct ClientIp | project nodeId = ClientIp, nodeType = "IP"),
    (edges | distinct Hostname | project nodeId = Hostname, nodeType = "Host");
edges
| make-graph ClientIp --> Hostname with nodes on nodeId
| graph-shortest-paths (src)-[e*1..5]->(dest)
    where src.nodeType == "IP" and dest.nodeId == "MAIL-SERVER01"
    project
        SourceIP = src.nodeId,
        PathLength = array_length(e),
        Hops = e.Hostname

Example: IRQL selectors -> make-graph -> connected components

Find clusters of IPs and domains that are interconnected -- potential C2 infrastructure:

let dns = Get_Dns_All;
let edges = dns | project Source = ClientIp, Target = Domain;
let nodes = union
    (edges | distinct Source | project nodeId = Source, nodeType = "IP"),
    (edges | distinct Target | project nodeId = Target, nodeType = "Domain");
edges
| make-graph Source --> Target with nodes on nodeId
| graph-mark-components with_component_id = ComponentId
| graph-to-table nodes
| summarize
    IPs = make_set_if(nodeId, nodeType == "IP"),
    Domains = make_set_if(nodeId, nodeType == "Domain"),
    Size = count()
  by ComponentId
| where Size > 3
| order by Size desc

Example: IRQL + make-graph integration

See references/EXAMPLES.md for multi-source investigation graphs combining IRQL selectors with make-graph, and Lift_To_Graph visual graph examples.

Practical Usage Scenarios

See references/SCENARIOS.md for full worked examples including:

• Reachability analysis (shortest paths to critical assets)
• Network segmentation validation (connected components)
• Blast radius of compromised accounts (variable-length path matching)
• Persistent graph models for SOC teams (graph_model + snapshots)

MCP Tools Used

Opening Queries in Kusto Explorer (Windows Only)

Optional convenience feature. The default workflow is to output the KQL in chat and let the user copy it into Kusto Explorer or the VS Code Kusto extension manually. Auto-launch is opt-in only.

Default: Output KQL in Chat

Always output the complete KQL with Step 1 (connect) and Step 2 (query) clearly labeled:

// Step 1: Connect to your cluster (skip if already connected)
// Example: uncomment to connect to the KC7 training cluster
// #connect cluster('kc7001.eastus.kusto.windows.net').database('ValdyTimes')
// Or replace with your own cluster:
// #connect cluster('<YOUR_CLUSTER>').database('<YOUR_DATABASE>')

// Step 2: Run the query below
<KQL_QUERY ending at make-graph>

Then immediately below, output an ADX Web Explorer version that appends | graph-to-table nodes as N, edges as E since ADX Web Explorer cannot render make-graph directly:

// ADX Web Explorer version (tabular output):
<SAME_QUERY>
| graph-to-table nodes as N, edges as E

This ensures the output works in both Kusto Explorer (graph visualization) and ADX Web Explorer (tabular results) without the user having to modify anything.

Optional: Save and Launch

If the user asks to save or open the query in Kusto Explorer, follow the procedure in references/KUSTO_EXPLORER_LAUNCH.md. Key rules:

• Always use ask_user to confirm before writing files or launching executables
• Always display the file contents in chat so the user can review before opening
• Never use shell interpolation or here-strings — write files via Set-Content/Add-Content
• Never encode queries into browser URLs
• On macOS/Linux, save the .kql file and suggest the VS Code Kusto extension or ADX Web Explorer

For make-graph visualization (the graph window), the query must end at make-graph — do not pipe to graph-match. Kusto Explorer only opens the graph visualization window when the output is a graph object, not a table.