当前位置：首页 > ds >正文

spring-ai-alibaba 之 graph 槽点

ds 2025/8/4 18:34:25

函数式编程

spring-ai-alibaba-graph-core 的代码也学习有一段时间了，有些地方确实还不尽如人意

其中有些底层设计感觉确实不太符合java开发人员的习惯，不太容易理解

究其根源，个人感觉应该是照搬 langgraph 的逻辑，强行使用函数式编程，没有进行java面向对象的适配导致的（反正我是不承认我不习惯大范围使用 java 的函数式编程的：））

毕竟在 langgraph 的官网介绍中都说了，Nodes 和 Edges 仅仅是 Python 函数

但做 Java 开发从没听过“仅仅是 Java Lambda表达式”或者“仅仅是Java Function”的说法，类比应该是“仅仅是 Java 类”或“仅仅是 Java 接口”

举个例子

原先 langgraph 的使用逻辑可能类似下面这样

def chatbot(state: State):return {"messages": [llm.invoke(state["messages"])]}llmNode = Node(chatbot)
graph_builder.add_node("chatbot", llmNode)

在python中，函数本身也是一种对象，可以当做参数进行传递

但是在java中，如果要完全复制这种操作，就需要引入Function等函数式编程，进行一层封装

NodeAction llmAction = overallState -> {Map<String, Object> map = ...//业务逻辑return map;
}Node llmNode = new Node(llmAction)

Node及NodeAction的定义如下，其中NodeAction是一个继承了Function接口的接口

public class Node {@FunctionalInterfacepublic interface NodeAction {Map<String, Object> apply(OverAllState state) throws Exception;}private final String id;private final NodeAction actionFactory;
}

但是这不够，在python中一个函数就能将参数转为异步

def chatbot(state: State):return {"messages": [llm.invoke(state["messages"])]}llmNode = Node(chatbot)
graph_builder.add_node("chatbot", async(llmNode))

那怎么办？spring-ai-alibaba-graph-core 为了支持这种写法，在外面又套了一层

public class Node {@FunctionalInterfacepublic interface AsyncNodeAction extends Function<OverAllState, CompletableFuture<Map<String, Object>>> {CompletableFuture<Map<String, Object>> apply(OverAllState state);@FunctionalInterfacepublic interface NodeAction {Map<String, Object> apply(OverAllState state) throws Exception;}static AsyncNodeAction node_async(NodeAction syncAction);}private final String id;private final AsyncNodeAction asyncActionFactory;
}

然后，python还支持多参数，即节点函数的入参包括State和RunnableConfig

def chatbot(state: State, runnableConfig: RunnableConfig):return {"messages": [llm.invoke(state["messages"])]}llmNode = Node(chatbot)
graph_builder.add_node("chatbot", async(llmNode))

那就在AsyncNodeAction外面再套一层AsyncNodeActionWithConfig，继承BiFunction

具体代码我就不贴了，后面还又套了一层ActionFactory

总之，让习惯java写法的人非常不习惯

如果是java开发一般会怎么写？一个接口搞定

public interface GraphNode {Map<String, Object> action(OverAllState state);default Map<String, Object> action(OverAllState state, RunnableConfig config) {return action(state);}default CompletableFuture<Map<String, Object>> asyncAction(OverAllState state) {return CompletableFuture.supplyAsync(() ->  action(state));}default CompletableFuture<Map<String, Object>> asyncAction(OverAllState state, RunnableConfig config) {return CompletableFuture.supplyAsync(() ->  action(state, config));}}

甚至普通节点都不需要异步，异步放到并行节点里去就ok了

public interface GraphNode {Map<String, Object> action(OverAllState state);default Map<String, Object> action(OverAllState state, RunnableConfig config) {return action(state);}}public interface ParallelNode extends GraphNode {default CompletableFuture<Map<String, Object>> asyncAction(OverAllState state) {return CompletableFuture.supplyAsync(() ->  action(state));}default CompletableFuture<Map<String, Object>> asyncAction(OverAllState state, RunnableConfig config) {return CompletableFuture.supplyAsync(() ->  action(state, config));}}

边其实也存在同样的问题，要封装为函数式的EdgeAction来操作