Class StateGraph<SD, S, U, N, I, O, C>

A graph whose nodes communicate by reading and writing to a shared state. Each node takes a defined State as input and returns a Partial<State>.

Each state key can optionally be annotated with a reducer function that will be used to aggregate the values of that key received from multiple nodes. The signature of a reducer function is (left: Value, right: UpdateValue) => Value.

See Annotation for more on defining state.

After adding nodes and edges to your graph, you must call .compile() on it before you can use it.

Example

import {
  type BaseMessage,
  AIMessage,
  HumanMessage,
} from "@langchain/core/messages";
import { StateGraph, Annotation } from "@langchain/langgraph";

// Define a state with a single key named "messages" that will
// combine a returned BaseMessage or arrays of BaseMessages
const StateAnnotation = Annotation.Root({
  sentiment: Annotation<string>,
  messages: Annotation<BaseMessage[]>({
    reducer: (left: BaseMessage[], right: BaseMessage | BaseMessage[]) => {
      if (Array.isArray(right)) {
        return left.concat(right);
      }
      return left.concat([right]);
    },
    default: () => [],
  }),
});

const graphBuilder = new StateGraph(StateAnnotation);

// A node in the graph that returns an object with a "messages" key
// will update the state by combining the existing value with the returned one.
const myNode = (state: typeof StateAnnotation.State) => {
  return {
    messages: [new AIMessage("Some new response")],
    sentiment: "positive",
  };
};

const graph = graphBuilder
  .addNode("myNode", myNode)
  .addEdge("__start__", "myNode")
  .addEdge("myNode", "__end__")
  .compile();

await graph.invoke({ messages: [new HumanMessage("how are you?")] });

// {
//   messages: [HumanMessage("how are you?"), AIMessage("Some new response")],
//   sentiment: "positive",
// }
Copy