itwanger/PaiAgent

🔥轻量级的AI工作流编排系统，类似dify、n8n，全程使用Vibe Coding，AI工具为Qoder+CLI。涉及到的技术栈包括SpringAI、LangGraph4J、SSE、MinIO、DAG自定引擎等。

PaiAgent – Enterprise‑grade visual AI workflow orchestration platform

What it is

PaiAgent is a self‑hosted, full‑stack platform that lets you build, schedule and run AI‑powered pipelines without writing code. A drag‑and‑drop editor (ReactFlow) on the front‑end lets business users and developers compose nodes that call large language models (OpenAI, DeepSeek, Alibaba Qwen, ZhiPu, etc.) and auxiliary tools (TTS, data I/O, custom plugins). At runtime the platform executes the graph either with a classic DAG engine (topological sort + cycle detection) or with a LangGraph4j state‑graph engine for richer conditional routing.

Core capabilities

Feature	What you get
Zero‑code workflow design	Visual canvas, node palette, parameter panels – no Java/TS required to assemble a pipeline
Dual execution engine	DAG for simple linear/branching flows; LangGraph4j for state‑graph, conditional and dynamic routing – selectable per workflow
Unified LLM access	Spring AI + Spring AI Alibaba adapters expose OpenAI‑compatible, DeepSeek, Qwen, ZhiPu and AIPing models through a common `ChatClient` API
Skills system	Declarative “skill” definitions in `SKILL.md` (YAML front‑matter + Markdown). Skills are auto‑registered, lazily loaded (summary → detail → references) and exposed to LLMs as function callbacks so a model can invoke them on its own
Extensible node model	Implement `NodeExecutor` (or extend `AbstractLLMNodeExecutor`) and register it; the platform will expose the node in the UI automatically
Real‑time debugging	Server‑Sent Events stream execution logs to a side‑drawer, showing start/success/error per node
Persistence & monitoring	MySQL stores workflow definitions and run logs; optional MinIO for file blobs; execution history view in the UI
Built‑in tool nodes	Input/Output, TTS (Qwen3 TTS, StepAudio), plus a plug‑in hook for any custom service

Typical use cases

Domain	Example workflow
Content creation	Generate article drafts → translate → polish → synthesize audio narration
Customer support	Capture user query → route to appropriate LLM (knowledge base) → generate response → log ticket
Data processing	Pull raw text → LLM‑based entity extraction → store results in a database
Media production	Script generation → TTS → embed audio in video metadata
Automation	Periodic report generation → email dispatch → archive PDF

Architecture at a glance

Frontend (React 18 + TypeScript)
  └─ ReactFlow canvas, Ant Design UI, Zustand state

Backend (Spring Boot 3.4, Java 21)
  ├─ Controllers → Services → MyBatis‑Plus DAOs
  ├─ Engine layer
  │   ├─ EngineSelector (dag ↔ langgraph)
  │   ├─ DAGParser (Kahn topological sort, DFS cycle check)
  │   └─ LangGraphWorkflowEngine (StateGraph, async node actions)
  ├─ Skill subsystem (SkillRegistry, progressive loading, FunctionCallback)
  └─ AI model layer (Spring AI + Spring AI Alibaba, ChatClientFactory)

Data layer
  ├─ MySQL (workflow definitions, execution logs)
  └─ MinIO (optional blob storage)

Technology stack

Layer	Tech	Version
Frontend	React, TypeScript, Vite, ReactFlow, Ant Design, Tailwind, Zustand	18.x / 5.x / latest
Backend	Spring Boot, Java 21, MyBatis‑Plus, FastJSON2, Maven	3.4.1 / 21+
AI integration	Spring AI (1.0.0‑M5), Spring AI Alibaba (1.0.0‑M6.1)	–
Workflow engines	Custom DAG engine, LangGraph4j Core (1.1.5) + Spring AI bridge (1.8.0‑beta3)	–
Storage	MySQL 8+, optional MinIO	–

Getting started (quick‑start)

Clone

git clone https://github.com/itwanger/PaiAgent-one.git
cd PaiAgent-one

Create MySQL database and run schema.sql (found under backend/src/main/resources).
Configure environment – copy .env.example in backend/ and frontend/ to .env / .env.local and fill in DB credentials, JWT secret, MinIO keys, etc.

Run the backend

cd backend
./mvnw spring-boot:run   # starts on http://localhost:8084

Run the frontend

cd frontend
npm install
npm run dev               # opens http://localhost:5173

Log in with the default dev account (admin / admin123) or set your own via env vars.
Create a workflow – drag an Input node, an OpenAI node, a TTS node and an Output node, connect them, configure the model API key, and hit Debug to see live SSE logs.

Roadmap highlights

v1.1 – conditional IF/ELSE nodes, loop (FOR/WHILE) nodes, more LLM providers (Claude, Gemini, local models), extra built‑in skills.
v1.5 – sub‑workflow calls, cron‑based scheduling, versioned workflows, template marketplace, image/video processing nodes.
v2.0 – multi‑tenant RBAC, collaborative editing, performance monitoring & alerts, i18n, third‑party plugin marketplace.

Who might benefit?

Product teams building AI‑augmented features without deep ML code.
Enterprises needing a controlled, auditable pipeline for LLM‑driven automation.
Developers who want a plug‑and‑play engine to prototype complex LLM orchestration (e.g., RAG, tool‑calling, multi‑model routing).

All information above is taken directly from the repository’s README; no external assumptions have been added.