Community

Blog Events Webinars Tutorials Forum
  1. Blog
  2. Events
  3. Webinars
  4. Tutorials
  5. Forum
Create Account
×
Community Blog Achieve Manus in A Dozen Lines of Code: A Quick Preview of Spring AI Alibaba Graph

Achieve Manus in A Dozen Lines of Code: A Quick Preview of Spring AI Alibaba Graph

This article shows you how to use Spring AI Alibaba to develop workflows and agent applications through three practical examples.

By Jun Liu

Based on Spring AI Alibaba Graph, developers can effortlessly develop workflows and intelligent agents & multiple agents in various modes, offering flexibility and richer features on the basis of Spring AI ChatClient.

In this article, we will show you how to use Spring AI Alibaba to develop workflows and agent applications through three practical examples, implementing intelligent agent orchestration in just a few lines of code.

  1. Example 1: A customer review processing system (implemented via workflow orchestration)
  2. Example 2: A weather forecast query system based on the ReAct Agent mode
  3. Example 3: OpenManus implementation based on Supervisor multi-agent system

For the complete source code of the Spring AI Alibaba Graph kernel and examples, see: https://github.com/alibaba/spring-ai-alibaba/tree/main/spring-ai-alibaba-graph

Example 1: A Customer Review Processing System Based on Workflow Orchestration

System architecture:

1

This example demonstrates a customer review processing system that first receives user comments and automatically classifies them based on the content of the comments, with a total of two levels of issue types.

  1. The first-level classification node divides comments into two types: positive and negative. If it is a positive comment, the workflow ends after the system records it; if it is a negative comment, it proceeds to the second-level classification.
  2. The second-level classification node identifies the user's specific problems in negative comments, such as "after-sale service", "product quality", and "transportation", and diverts them to specific processing nodes according to specific problems.
  3. After the problem is processed and recorded, the workflow terminates.

Core code snippet:

AgentStateFactory<OverAllState> stateFactory = (inputs) -> {
            OverAllState state = new OverAllState();
            state.registerKeyAndStrategy("input", new ReplaceStrategy());
            state.registerKeyAndStrategy("classifier_output", new ReplaceStrategy());
            state.registerKeyAndStrategy("solution", new ReplaceStrategy());
            state.input(inputs);
            return state;
        };

StateGraph stateGraph = new StateGraph("Consumer Service Workflow Demo", stateFactory)
            .addNode("feedback_classifier", node_async(feedbackClassifier))
            .addNode("specific_question_classifier", node_async(specificQuestionClassifier))
            .addNode("recorder", node_async(new RecordingNode()))

            .addEdge(START, "feedback_classifier")
            .addConditionalEdges("feedback_classifier",
                    edge_async(new CustomerServiceController.FeedbackQuestionDispatcher()),
                    Map.of("positive", "recorder", "negative", "specific_question_classifier"))
            .addConditionalEdges("specific_question_classifier",
                    edge_async(new CustomerServiceController.SpecificQuestionDispatcher()),
                    Map.of("after-sale", "recorder", "transportation", "recorder", "quality", "recorder", "others",
                            "recorder"))
            .addEdge("recorder", END);

You can download the source code of this example and run it. Open a browser to visit the following example links to view the running effect:

• http://localhost:18080/customer/chat?query=The product I received was damaged during delivery. How can I get a return or exchange?

• http://localhost:18080/customer/chat?query=My product is not working properly. How can I get it repaired?

• http://localhost:18080/customer/chat?query=I received the product - it's great. I will definitely buy it again.

2. Example 2: A Weather Forecast Query System Based on the ReAct Agent Mode

The following figure shows the React Agent architecture:

2

In this example, only one weather query service is bound to the agent. After receiving a weather query service from the user, the process loops between AgentNode and ToolNode until the user's command is completed. The condition for determining the completion of the command (that is, the ReAct end condition) is also simple. The process ends when the model's AssistantMessage contains no tool_call command (using the default behavior).

ReactAgent reactAgent = ReactAgent.builder()
    .name("React Agent Demo")
    .prompt("Please proceed to complete the task that the user enters for you.")
    .chatClient(chatClient)
    .resolver(resolver)
    .maxIterations(10)
    .build();

reactAgent.invoke(Map.of("messages", new UserMessage(query)))

You can download the source code of this example and run it. Open a browser to visit the following example link to view the running effect:

• http://localhost:18080/react/chat?query=Help me check the weather in Hangzhou, Shanghai, and Nanjing respectively.

3. Example 3: OpenManus Implementation Based on Supervisor Multi-agent System

Spring AI Alibaba released the industry's first Java implementation solution for OpenManus. For the source code of the original implementation and blog interpretation articles, see:

  1. Blog interpretation: https://java2ai.com/blog/spring-ai-alibaba-openmanus
  2. Sample source code: https://github.com/alibaba/spring-ai-alibaba/community/openmanus

The original OpenManus implementation did not use Spring AI Alibaba Graph, requiring us to invest significant effort in writing workflow control logic. In previous versions of OpenManus, we summarized the following implementation-related issues:

• 80% of the code in the repository was dedicated to workflow orchestration, including connecting manus agent sub-processes, memorizing messages, calling forwarding tools, and modifying the global state. These tasks could be handled by a highly abstracted agent framework to simplify development complexity.

• The coverage and execution effect of tools were suboptimal, particularly for functionalities like browser operations and script execution tools.

• In the planning and workflow processes, you cannot perform manual reviews, dynamic modifications, rollbacks, and other actions.

• It is relatively difficult to debug the current OpenManus implementation.

Now, with the help of Spring AI Alibaba Graph, we can implement the same OpenManus system with 80% less code than before. The following is the architecture diagram of OpenManus implementation:

3

In the OpenManus example, we have implemented a multi-agent system where three core agents collaborate to accomplish user tasks:

  1. Planning Agent, which is responsible for task planning and generating detailed step-by-step execution plans.
  2. Supervisor Agent, which oversees the Executor Agent in completing tasks planned by the Planning Agent, forwarding each subtask to the Executor Agent in sequence.
  3. Executor Agent, which handles subtask execution with multiple available tools such as Browser_use, FileSaver, and PythonExecuter.

You can download the source code of this example and run it. Open a browser to visit the following example link to view the running effect:

• http://localhost:18080/manus/chat?query=Help me check Alibaba's stock information in the past week.

4. Run the Example

Run the following command to download the source code:

git clone https://github.com/alibaba/spring-ai-alibaba.git
cd spring-ai-alibaba-graph/spring-ai-alibaba-graph-example

To access the LLM, you need to export the model API key as an environment variable:

export AI_DASHSCOPE_API_KEY=xxx

You can then start the sample application by running the GraphApplication class directly in the IDE.

Alternatively, you can run the following Maven command to enable the sample application (note that since Spring AI Alibaba Graph hasn't been officially released yet, you need to install the source code in the root directory first):

mvn clean install
cd spring-ai-alibaba-graph/spring-ai-alibaba-graph-example
mvn spring-boot:run

5. Design Philosophy and Future Planning

The current version of Spring AI Alibaba Graph draws significant design inspiration from Langgraph, including global state management and graph definition. It can essentially be considered to be a Java implementation of Langgraph. Special thanks to the Langchain community for their open-source Langgraph intelligent agent framework, allowing us to move forward upon the achievements of predecessors. As Agent technology is evolving rapidly, we will explore more cutting-edge implementation solutions on this basis.

Spring AI Alibaba Graph has not been officially released. Developers are welcome to experience and contribute to the project through the GitHub source code: https://github.com/alibaba/spring-ai-alibaba/tree/main/spring-ai-alibaba-graph

AI Artificial Intelligence Open Source Generative AI GenAI Spring AI Spring AI Alibaba Graph ReAct Agent OpenManus
0 1 0
Share on

Read previous post:

Break Through Data Silos: Practices of Multi-cloud Observability Integration Based on Object Storage Service (OSS)

Read next post:

Non-intrusive Go Application Observability Based on Docker Multi-stage Builds

Alibaba Cloud Native Community

634 posts | 55 followers

You may also like

Comments

Alibaba Cloud Native Community

634 posts | 55 followers

Related Products

More Posts by Alibaba Cloud Native Community

See All