Skip to content

Migrating

Migrating from plain Semantic Kernel

This guide shows how to move from ad-hoc pipelines built with Semantic Kernel (SK) to structured graphs using SemanticKernel.Graph with minimal changes.

Prerequisites

  • SK configured with your LLM provider
  • Existing plugins/functions you want to reuse

1) Install and configure

var builder = Kernel.CreateBuilder();
// ... your SK config (models, memory, logging)
// Add graph support
builder.AddGraphSupport(options =>
{
    options.EnableLogging = true;
});
var kernel = builder.Build();

2) Wrap existing functions as nodes

// From plugin function
var myNode = FunctionGraphNode.FromFunction(myPlugin["MyFunction"], nodeId: "my-node");

// REST tool → node
var restNode = new RestToolGraphNode(new RestToolSchema { /* map inputs/outputs */ });

3) Connect nodes into a graph

// Create executor and add nodes
var graph = new GraphExecutor("my-graph");
graph.AddNode(myNode);
graph.AddNode(restNode);

// Connect by node ids (recommended) or use Connect overloads that accept node instances
graph.Connect(myNode.NodeId, restNode.NodeId);

4) Add control-flow when needed

var decision = new ConditionalGraphNode(state => (bool)state["should_call_api"]);
graph.AddNode(decision);

// Wire conditional paths explicitly using AddTrueNode/AddFalseNode
decision.AddTrueNode(restNode);
decision.AddFalseNode(myNode);

// Connect decision into the main flow
graph.Connect(myNode.NodeId, decision.NodeId);

5) Execute with shared state

var args = new KernelArguments { ["input"] = userMessage };
var result = await graph.ExecuteAsync(kernel, args);

Feature mapping

  • SK function pipeline → FunctionGraphNode chain
  • Conditional branching → ConditionalGraphNode/SwitchGraphNode
  • Loops → WhileLoopGraphNode/ForeachLoopGraphNode
  • Error handling/retry → ErrorHandlerGraphNode/RetryPolicyGraphNode
  • Observability → GraphPerformanceMetrics + IGraphTelemetry
  • Checkpoint/recovery → CheckpointingExtensions

Migration checklist

  • Reuse plugins/functions via FunctionGraphNode
  • External APIs via RestToolGraphNode + IToolSchemaConverter
  • Replace custom flow code with nodes/edges
  • Enable metrics/telemetry and (optionally) checkpointing
  • Add streaming if you need real-time events