Switching Spring AI from DashScope to Ollama: Multi‑MCP Server Calls and Targeted Server Example

As large language model (LLM) usage expands, developers often want to move projects built on commercial cloud APIs such as Alibaba Cloud DashScope to locally hosted Ollama models for lower cost, better privacy, and offline development.

Spring AI model abstraction layer

Spring AI defines a unified ChatModel interface for different model providers. Switching models only requires changing the starter dependency and configuration; the business logic remains untouched.

Advantages of Ollama

Free and runs locally – no API key needed.

Provides an OpenAI‑compatible HTTP API for easy integration.

Supports a rich ecosystem of open‑source models (e.g., Qwen, Llama, Mistral) that can be pulled with a single command.

Implementation steps

1. Modify pom.xml dependencies

<parent>
  <groupId>org.springframework.boot</groupId>
  <artifactId>spring-boot-starter-parent</artifactId>
  <version>3.4.5</version>
</parent>
<groupId>com.badao</groupId>
<artifactId>spring-ai-ollama-mcp-third-multipart</artifactId>
<version>1.0.0</version>
<properties>
  <java.version>17</java.version>
  <maven.compiler.release>17</maven.compiler.release>
  <spring-ai.version>1.1.2</spring-ai.version>
</properties>
<dependencyManagement>
  <dependencies>
    <dependency>
      <groupId>org.springframework.ai</groupId>
      <artifactId>spring-ai-bom</artifactId>
      <version>${spring-ai.version}</version>
    </dependency>
    <!-- Spring AI MCP client -->
    <dependency>
      <groupId>org.springframework.ai</groupId>
      <artifactId>spring-ai-starter-mcp-client</artifactId>
    </dependency>
  </dependencies>
</dependencyManagement>
<repositories>
  <repository>
    <id>spring-milestones</id>
    <name>Spring Milestones</name>
    <url>https://repo.spring.io/milestone</url>
  </repository>
  <repository>
    <id>aliyun</id>
    <name>Aliyun Maven</name>
    <url>https://maven.aliyun.com/repository/public</url>
  </repository>
</repositories>

2. Adjust application.yml configuration

server:
  port: 8082
spring:
  ai:
    ollama:
      base-url: http://localhost:11434   # Ollama service address
      chat:
        model: qwen2.5:7b-instruct   # model that supports tool calling
        options:
          temperature: 0.7
    mcp:
      client:
        enabled: true
        request-timeout: 60s
        toolcallback:
          enabled: true
        streamable-http:
          connections:
            hacker-news:
              url: https://hn.caseyjhand.com
            nws-weather:
              url: https://nws.caseyjhand.com/mcp
logging:
  level:
    org.springframework.ai.mcp: DEBUG

3. Zero code changes

All Java classes (e.g., AiConfig, AiService, AiController) remain unchanged. Spring AI injects the active ChatModel via ChatClient.Builder, fully decoupling business logic from the underlying model implementation.

Test results

Two queries were sent:

{
  "message": "Answer two questions: what is the hottest Hacker News article and what is the weather in New York?"
}

The response is shown in the image below.

A single‑question query was also tested (image omitted for brevity).

Common issues and troubleshooting

DashScope API key must be set : Residual DashScope starter triggers auto‑configuration. Remove the starter and run mvn clean before restarting.

Missing dependency version : Some Spring AI starters are not fully managed by the BOM. Explicitly set ${spring-ai.version} for those artifacts.

Cannot find spring-ai-ollama-spring-boot-starter:1.1.2 : The correct artifactId in version 1.1.2 is spring-ai-starter-model-ollama. Update the dependency accordingly.

Connection reset after startup : HTTP/2 or SSE streams may be reset by network devices. Increase request-timeout to 60s and optionally set per‑connection timeouts. The framework will auto‑reconnect; the warning can be ignored or logged at a higher level.

Model does not support tool calling : Some Ollama models (e.g., early Llama 2) lack Function Calling. Switch to a model that explicitly supports it, such as qwen2.5:7b, llama3.1:8b, or mistral:7b.

Key takeaways

Dependency isolation: keep only one model starter in a Spring AI project to avoid conflicting auto‑configurations.

Configuration‑driven model switching: change the model by editing application.yml without touching business code.

Use the Spring AI BOM for version management; manually specify versions for any unmanaged artifacts.

Network resilience: the MCP client automatically retries; occasional connection resets are harmless.

Model capability matters: ensure the selected Ollama model supports Function Calling before relying on tool usage.