Skip to content

workflows.md

Latest commit

 

History

History
97 lines (71 loc) · 4.34 KB

workflows.md

File metadata and controls

97 lines (71 loc) · 4.34 KB
Raw

Workflows

The fluent Workflow SDK creates typed Conductor workflow definitions in Java. Start with the maintained workflow and worker Hello World; it registers a task definition, starts a worker, runs a workflow, and prints Result: PASSED.

Prerequisites

SIMPLE tasks require both a registered task definition and a worker polling the exact task name. The Hello World example is the shortest complete reference for that lifecycle.

Build a workflow

Fragment — create this after constructing a WorkflowExecutor for your server.

import com.netflix.conductor.common.metadata.workflow.WorkflowDef;
import com.netflix.conductor.sdk.workflow.def.ConductorWorkflow;
import com.netflix.conductor.sdk.workflow.def.WorkflowBuilder;
import com.netflix.conductor.sdk.workflow.def.tasks.SimpleTask;
import com.netflix.conductor.sdk.workflow.executor.WorkflowExecutor;

WorkflowExecutor executor = new WorkflowExecutor("http://localhost:8080/api");

ConductorWorkflow<java.util.Map<String, Object>> workflow =
        new WorkflowBuilder<java.util.Map<String, Object>>(executor)
                .name("quote_workflow")
                .version(1)
                .description("Calculate and email an insurance quote")
                .ownerEmail("team@example.com")
                .timeoutPolicy(WorkflowDef.TimeoutPolicy.TIME_OUT_WF, 300)
                .add(
                        new SimpleTask("calculate_quote", "calculate_quote"),
                        new SimpleTask("send_email", "send_email")
                                .input("email", "${workflow.input.email}")
                                .input("subject", "Your quote is ${calculate_quote.output.amount}"))
                .build();

ConductorWorkflow can register a reusable definition, execute an already registered definition, or submit a one-off dynamic definition:

workflow.registerWorkflow(true, true);             // overwrite and register missing task definitions
workflow.execute(java.util.Map.of("email", "user@example.com"));
workflow.executeDynamic(java.util.Map.of("email", "user@example.com"));

Use dynamic execution only when each run truly needs a distinct graph; registered definitions are easier to inspect, reuse, and evolve.

System tasks

The builder accepts all task classes in com.netflix.conductor.sdk.workflow.def.tasks, including ForkJoin, Wait, Switch, DynamicFork, DoWhile, Join, Dynamic, Terminate, SubWorkflow, and SetVariable.

Read the current source when you need a task-specific builder: workflow task classes.

AI and vector tasks

The fluent SDK exposes current Conductor system-task types. Provider and vector database integrations are configured on the server.

Fragment — index text:

import org.conductoross.conductor.sdk.ai.LlmIndexText;

LlmIndexText indexText = new LlmIndexText("index_doc", "index_ref")
        .vectorDb("knowledge_base")
        .namespace("documents")
        .index("articles")
        .embeddingModelProvider("openai")
        .embeddingModel("text-embedding-3-small")
        .text("${workflow.input.document}")
        .docId("${workflow.input.documentId}");

Fragment — search the index:

import org.conductoross.conductor.sdk.ai.LlmSearchIndex;

LlmSearchIndex search = new LlmSearchIndex("search_docs", "search_ref")
        .vectorDb("knowledge_base")
        .namespace("documents")
        .index("articles")
        .embeddingModelProvider("openai")
        .embeddingModel("text-embedding-3-small")
        .query("${workflow.input.question}")
        .maxResults(5);

Add either task with .add(indexText) or .add(search). The exact embedding model must match the model used to index the data.

Related guides

  • Workers — implement and run SIMPLE tasks.
  • Workflow test harness — test definitions and workers against a local server.
  • Agent SDK — build durable LLM agents, tools, and dynamic plans.