Spring AI Alibaba

Spring AI Alibaba Adapter

AgentRuntime implementation backed by Spring AI Alibaba (Alibaba Cloud AI) ReactAgent. Spring AI Alibaba extends Spring AI with multi-agent orchestration patterns (SequentialAgent, ParallelAgent, RoutingAgent, LoopAgent) and a graph runtime; this adapter exposes them through @AiEndpoint.

Maven Coordinates

<dependency>
    <groupId>org.atmosphere</groupId>
    <artifactId>atmosphere-spring-ai-alibaba</artifactId>
    <version>${project.version}</version>
</dependency>
<dependency>
    <groupId>com.alibaba.cloud.ai</groupId>
    <artifactId>spring-ai-alibaba-agent-framework</artifactId>
    <version>${spring-ai-alibaba.version}</version>
</dependency>
<dependency>
    <groupId>org.springframework.ai</groupId>
    <artifactId>spring-ai-starter-model-openai</artifactId>
    <version>${spring-ai-alibaba-spring-ai.version}</version>
</dependency>

How It Works

Drop the dependency alongside atmosphere-ai and the framework auto-detects it via ServiceLoader. AtmosphereSpringAiAlibabaAutoConfiguration wires the ReactAgent into the SPI when running under Spring Boot:

@AiEndpoint(path = "/ai/chat", systemPrompt = "You are a helpful assistant")
public class MyChat {

    @Prompt
    public void onPrompt(String message, StreamingSession session) {
        session.stream(message);  // uses Spring AI Alibaba automatically
    }
}

Buffered Streaming — Important

Spring AI Alibaba’s ReactAgent.call(List<Message>) → AssistantMessage is synchronous as of v1.1.2.2; there is no native Flux<...> streaming agent method. This adapter wraps call(...) and delivers the full reply as a single session.send(text) chunk followed by session.complete(). The Atmosphere transport still frames that chunk as a streamed message, so the UI sees a streaming response — but there are no incremental token deltas from the LLM.

If token-by-token streaming matters, drive Spring AI’s StreamingChatModel directly via atmosphere-spring-ai. You will lose Spring AI Alibaba’s multi-agent / graph orchestration patterns; that’s the trade.

Requirements — Spring Boot 3 only today

Spring AI Alibaba 1.1.2.2 is compiled against Spring AI 1.1.6, which pins the Spring Boot 3-era FQN of RestClientAutoConfiguration (Spring Boot 4 ships it at a renamed FQN). Run this adapter under a Spring Boot 3 application:

cd samples/spring-boot-ai-chat
./mvnw -Pspring-boot3 spring-boot:run

A Spring Boot 4 path becomes possible once Alibaba publishes a Spring AI 2.x-aligned spring-ai-alibaba-agent-framework. The sibling atmosphere-agentscope adapter is unaffected and works on Spring Boot 4.

Per-Request Extension

SpringAiAlibabaRunnableConfig exposes Alibaba’s per-invocation RunnableConfig — threadId (memory-thread continuation), checkPointId (resume), streamMode, metadata, and store. See the AI Reference for the per-request sidecar pattern.

Samples

Spring Boot AI Chat — run with -Pspring-boot3 and drop atmosphere-spring-ai-alibaba on the classpath