Loops

For many use cases, designing linear workflows might not be enough: you might need to add some quality checking, and re-execute certain steps based on the results of that, or you might want to refine the responses coming from an LLM or a RAG pipeline (e.g.) by iterating on the generation steps.

In this sense, looping (i.e. creating cyclic workflows) can be a very useful pattern to learn when building with LlamaIndex Workflows.

Since workflows are event-driven, and every step is triggered by an event, you can go back to a specific step at any point - you just need to emit the corresponding trigger event!

Here is how you can do it with code:

import { createWorkflow, workflowEvent } from "@llamaindex/workflow-core";
import { createStatefulMiddleware } from "@llamaindex/workflow-core/middleware/state";

type AgentWorkflowState = {
  counter: number,
  max_counter: number
};

const { withState } = createStatefulMiddleware(
  (state: AgentWorkflowState) => state,
);
const workflow = withState(createWorkflow());

const startEvent = workflowEvent<void>();
const increaseCounterEvent = workflowEvent<void>();
const stopEvent = workflowEvent<number>();

workflow.handle([startEvent], async (context, { data }) => {
    const { sendEvent, state } = context;
    if (state.counter < state.max_counter) {
        sendEvent(increaseCounterEvent.with())
    } else {
        sendEvent(stopEvent.with(state.counter))
    }
})

workflow.handle([increaseCounterEvent], async (context, { data }) => {
    const { sendEvent, state } = context;
    state.counter += 1
    sendEvent(startEvent.with())
})

const { stream, sendEvent } = workflow.createContext({
  counter: 0,
  max_counter: 5,
});

sendEvent(startEvent.with(),);

const result = await stream.until(stopEvent).toArray();

// should print 5 since the workflow is looping
console.log(result[result.length - 1].data)