🌟 Starring me: If you find this repository beneficial, don't forget to give it a star! 🌟 It's a simple way to show your appreciation and help this project grow!

Overview

JavAI Workflow (named initially Langchain4j-workflow) is a dynamic, stateful workflow engine crafted as a Java library. It empowers java developers with granular control over the orchestrated workflows as a graph, iteratively, with cycles, flexibility, control, and conditional decisions. This engine is a game-changer for building sophisticated AI applications, such as multiples RAG-based approaches using modern paradigms and agent architectures. It enables the crafting of custom behavior, leading to a significant reduction in hallucinations and an increase in response reliability.

jAI Workflow is influenced by LangFlow, LangGraph, Graphviz.

Anatomy

Node

A Node represents a single unit of work within the workflow. It encapsulates a specific function or task that processes the stateful bean and updates it. Nodes can be synchronous or asynchronous (streaming).

Module

A Module is a collection of nodes grouped together to perform a higher-level function. Modules can be reused across different workflows, providing modularity and reusability.

Workflow

A Workflow is a directed graph of nodes, modules and edges that defines the sequence of operations to be performed. It manages the state transitions and execution flow, ensuring that each node processes the stateful bean in the correct order.

Principles:

• Java-based jAI workflows are configuration as code (java) and agnostic to AI models, enabling you can define custom advanced and dynamic workflows just writing Java code.
• Stateful: jAI Workflow is a stateful engine, enabling you to design custom states as POJO and transitions. This feature provides a robust foundation for managing the flow and state of your application.
• Graph-Based: The workflow is graph-based, offering the flexibility to define custom workflows with multiple directions such as one-way, round trip, loop, recursive and more. This feature allows for intricate control over the flow of your application.
• Flexible: jAI Workflow is designed with flexibility in mind. You can define custom workflows, modules or agents to build RAG systems as LEGO-like. A module can be decoupled and integrated in any other workflow.
• AI Ecosystem integration: jAI workflow is designed to be potentially integrable with any emerging AI technology in the future for java.
• Publish as API: jAI Workflow can be published as an API as entrypoint once you have defined your custom workflow. This feature allows you to expose your workflow as a service.
• Observability: jAI Workflow provides observability features to monitor the execution of the workflow, trace inputs and outputs, and debug the flow of your application.
• Scalable: jAI Workflow can be deployed as any java project as standalone or distributed through containers in any cloud provider or kubernetes environment. Each module can run in a different JVM env or container for scalability in production environments.

🚀 Features

v0.3.0 Features

• Workflow lifecycle: Lifecycle is divided into 3 stages: definition, build, run. This feature allows you to define parameters to customize the workflow behavior for each stage, giving you the flexibility to redefine, rebuild, and rerun workflows at any time.
• Workflow patterns: This version supports multiple workflow patterns, including one-way, parallelization, routing, branching and conditional inspired on Modular RAG AI paper. All of these patterns can be used at definition, build and run time, except for parallelization which is not supported at run time.
• Node types: Adds new node types: Split, Merge, Parallel, Conditional with a list of expected nodes to be returned. These nodes allow you to define complex workflows patterns with multiple transitions between nodes.
• Debug workflow state: Once workflow runs you can get a list of ComputedTransition with all details of the workflow execution. This feature allows you to debug the flow of your application and trace the inputs, outputs, execution order, datetime of each node.
• Workflow JIT modification: You can put edges myworkflow.putEdge(..), add more nodes myworkflow.addNode(..) and override the start node myworkflow.startNode(..) at Just-in-time after workflow runs. This feature allows you to modify the workflow behavior dynamically during execution.
• Workflow visualization: You can generate the workflow image at definition time and at runtime. This feature allows you to visualize the transitions computed of your app workflow. Both kinds of images can be generated in a given path File or as BufferedImage to be displayed in a java notebook. Also, you can use StyleGraph.SKETCHY as StyleAttribute to generate workflow images with excalidraw style. This style is supported in Graphviz implementation only.

v0.2.0 Features

• Graph-core: The engine supports create Nodes, Conditional Nodes, Edges, and workflows as a graph. This feature allows you to define custom workflows with multiple Transitions between nodes such as one-way, round trip and recursive.
• Run workflow: jAI Workflow supports synchronized workflow.run() and streaming worflow.runStream() runs the outputs as they are produced by each node. This last feature allows for real-time processing and response in your application.
• Integration: LangChain4j integration, enabling you to define custom workflows using all the features that LangChain4j offers. This integration provides a comprehensive toolset for building advanced AI applications to integrate with multiple LLM providers and models.
• Visualization: The engine supports the generation of workflow images. This feature allows you to visualize the flow computed of your app workflow. By Default it uses Graphhviz lib to generate the image, but you implement your own image generator on GraphImageGenerator.java interface.

Q1 2025 Features

• Graph-Core:
  • [x] Split Nodes
  • [x] Merge Nodes
  • [x] Parallel transitions (except runtime)
  • Human-in-the-loop
• Modular (Group of nodes):
  • Module
  • Remote Module
• Integration:
  • Model Context Protocol (MCP) integration as server and client.
  • Define remote module as MCP server.
• API:
  • Publish workflow as API (SSE for streaming runs and REST for sync runs).

🗺️ Future Features

• Deployment Model:
  • Dockerize workflow
  • Kubernetes deployment
  • Cloud deployment
• Observability:
  • OpenTelemetry integration (metrics and traces).
  • Debugging mode logging structure.
• Playground:
  • Web-based playground to add, test and run jAI workflows APIs.
  • Chatbot Q&A viewer.
  • Graph tracing visualization for debugging
  

  This is a prototype, the final version will be available soon. Open an issue if you want to share your ideas or contribute to this feature.

Reference Architectures

jAI Workflow is designed with a modular architecture, enabling you to define custom workflows, modules, or agents to build RAG systems as LEGO-like. A module can be decoupled and integrated into any other workflow.

Standalone Architecture