当前位置：首页 > news >正文

从零构建轻量级 DAG 编排引擎：处理大模型复杂工作流的实战

news 2026/6/17 18:29:11

从零构建轻量级 DAG 编排引擎：处理大模型复杂工作流的实战

一、为什么简单的链式调用不够用

在真实业务里，单靠一个 Prompt 很难处理复杂的流程。开发者通常会把多个 LLM 调用、API 请求和数据清洗步骤串在一起。但一旦逻辑变复杂，这种线性调用就会出问题。

代码里开始出现大量的if-else嵌套，处理异步等待和节点依赖变得非常麻烦。这时候，有向无环图（DAG）是一个更稳妥的选择。它不仅能理清任务顺序，还能让没有依赖关系的节点并行执行，减少等待时间。

二、DAG 调度逻辑与拓扑排序

工作流里的每个步骤就是一个“节点”，节点间的依赖关系就是“边”。要执行这个图，核心是先做环路检测，再通过拓扑排序确定执行顺序。

下面的图展示了一个简单的流程：先清洗输入，然后并行做情感分类和关键词提取，最后汇总生成报告。

graph LR Start([启动]) --> NodeA[清洗输入] NodeA --> NodeB[情感分类] NodeA --> NodeC[关键词提取] NodeB --> NodeD[生成报告] NodeC --> NodeD NodeD --> End([结束]) style NodeB fill:#bbf,stroke:#333,stroke-width:2px style NodeD fill:#bfb,stroke:#333,stroke-width:2px

节点 B 和 C 都依赖 A，但它们之间没关系，所以引擎会让它们同时跑。

三、Node.js 轻量级实现

这是一个基于 JavaScript 的简单 DAG 引擎。它实现了拓扑排序来检查依赖，并支持异步并发执行。

class WorkflowNode { constructor(id, taskFunction) { this.id = id; this.taskFunction = taskFunction; this.dependencies = []; this.status = 'PENDING'; this.result = null; } addDependency(nodeId) { this.dependencies.push(nodeId); } } class DagEngine { constructor() { this.nodes = new Map(); } registerNode(node) { this.nodes.set(node.id, node); } // 拓扑排序：检查环路并决定顺序 resolveExecutionOrder() { const inDegree = new Map(); const adjList = new Map(); const order = []; for (const [id, node] of this.nodes) { inDegree.set(id, 0); adjList.set(id, []); } for (const [id, node] of this.nodes) { for (const depId of node.dependencies) { if (!this.nodes.has(depId)) { throw new Error(`节点 ${id} 依赖的 ${depId} 未注册`); } adjList.get(depId).push(id); inDegree.set(id, inDegree.get(id) + 1); } } const queue = []; for (const [id, degree] of inDegree) { if (degree === 0) queue.push(id); } while (queue.length > 0) { const currId = queue.shift(); order.push(currId); for (const nextId of adjList.get(currId)) { inDegree.set(nextId, inDegree.get(nextId) - 1); if (inDegree.get(nextId) === 0) { queue.push(nextId); } } } if (order.length !== this.nodes.size) { throw new Error("检测到循环依赖，无法执行"); } return order; } // 并发执行 async executeWorkflow(inputContext) { const completedResults = { ...inputContext }; const runningPromises = new Map(); while (true) { let hasPending = false; let progressed = false; for (const [id, node] of this.nodes) { if (node.status === 'COMPLETED' || node.status === 'FAILED') continue; hasPending = true; if (node.status === 'RUNNING') continue; // 检查依赖是否都完成了 const allDepsMet = node.dependencies.every(depId => { const depNode = this.nodes.get(depId); return depNode && depNode.status === 'COMPLETED'; }); if (allDepsMet) { node.status = 'RUNNING'; progressed = true; const promise = (async () => { try { const depData = {}; node.dependencies.forEach(depId => { depData[depId] = this.nodes.get(depId).result; }); node.result = await node.taskFunction(completedResults, depData); node.status = 'COMPLETED'; } catch (error) { node.status = 'FAILED'; throw error; } })(); runningPromises.set(id, promise); } } if (!hasPending) break; if (!progressed && runningPromises.size === 0) { throw new Error("死锁：没有节点能继续执行"); } await Promise.race(runningPromises.values()); for (const [id, promise] of runningPromises) { const node = this.nodes.get(id); if (node.status === 'COMPLETED' || node.status === 'FAILED') { runningPromises.delete(id); } } } const finalOutput = {}; for (const [id, node] of this.nodes) { finalOutput[id] = node.result; } return finalOutput; } } // 测试运行 (async () => { const engine = new DagEngine(); const nodeA = new WorkflowNode('CleanInput', async (context) => { return context.rawText.trim().replace(/[<>]/g, ''); }); const nodeB = new WorkflowNode('LlmClassify', async (context, depData) => { const text = depData.CleanInput; await new Promise(resolve => setTimeout(resolve, 500)); // 模拟 API 延迟 return text.includes("好") ? "POSITIVE" : "NEGATIVE"; }); nodeB.addDependency('CleanInput'); const nodeC = new WorkflowNode('ExtractKeywords', async (context, depData) => { const text = depData.CleanInput; return text.split(' ').filter(word => word.length > 1); }); nodeC.addDependency('CleanInput'); const nodeD = new WorkflowNode('GenerateReport', async (context, depData) => { const sentiment = depData.LlmClassify; const keywords = depData.ExtractKeywords; return `情感: ${sentiment}, 关键词: [${keywords.join(', ')}]`; }); nodeD.addDependency('LlmClassify'); nodeD.addDependency('ExtractKeywords'); engine.registerNode(nodeA); engine.registerNode(nodeB); engine.registerNode(nodeC); engine.registerNode(nodeD); const order = engine.resolveExecutionOrder(); console.log("执行顺序:", order.join(' -> ')); const result = await engine.executeWorkflow({ rawText: " 这个产品设计得非常 好，解决了我的痛点。 " }); console.log("结果:", result); })();