我爱 Spring AI。这是一个很棒的项目，旨在将 AI 工程的模式和实践带给 Spring Boot 开发者。它提供了简洁、惯用的抽象，让任何 Spring 开发者都能感到宾至如归，并且它还集成了各种不同的向量存储、嵌入模型、转录模型、图像模型和聊天模型。

新版本 m4，基于 Spring Boot 3.4 构建，并增加了大量新功能。一如既往，我无法涵盖所有新功能，但发布说明做得非常出色。

• 增加了对 Amazon Bedrock Converses 的新支持
• 在 Java 和 Kotlin 中，对更具表现力的函数调用进行了大量工作
• 这是对 AI 社区中关于“高级和模块化 RAG”的想法的首次实现。RAG，当然是检索增强生成，指的是使用系统或服务中的数据来告知聊天模型生成的响应。这些想法在本 篇论文和 这篇论文中有所阐述，而构建块正在这个新版本中生根发芽。太棒了！
• across the board to various vector store integrations and chat models
• there’saeven a comprehensive chat model comparison page in the documentation
• vector storage and embedding improvements, including enhancements to vector stores like Azure and Milvus

The evolving nature of the functional callback support has so enamored me. Spring AI aims to make connecting your AI models with your data and business logic easy. Remember: the name of the game here is integration. Most people aren’t going to build their models. They’re going to integrate existing ones into their business logic and services. And where does all that stuff live? In Spring, of course. Spring AI is a natural! And it keeps getting easier. In this release, there’s new support for describing and then letting models invoke functions if they decide they’ve got the need to do so.

Here’s a simple example demonstrating the definition of a FunctionCallback and the Spring AI ChatClient, which is your first port of call for all interactions with a Sprign AI ChatModel.

package com.example.bootiful_34.ai;

import org.springframework.ai.chat.client.ChatClient;
import org.springframework.ai.model.function.FunctionCallback;
import org.springframework.context.annotation.Bean;
import org.springframework.context.annotation.Configuration;

@Configuration
class AiConfiguration {

	@Bean
	ChatClient chatClient(ChatClient.Builder builder, FunctionCallback weatherFunctionCallback) {
		return builder.defaultFunctions(weatherFunctionCallback).build();
	}

	@Bean
	FunctionCallback weatherFunctionCallback() {
		return FunctionCallback.builder()
			.description("returns the weather for a given city")
			.function("getCurrentWeatherForACity",
					(WeatherRequest wr) -> new WeatherResponse(wr.city(),
							wr.city().equalsIgnoreCase("san francisco") ? 68.0f : 72.0f))
			.inputType(WeatherRequest.class)
			.build();
	}

}

record WeatherRequest(String city) {
}

record WeatherResponse(String city, float temperature) {
}

It’s a pretty trivial example: given a WeatherRequest specifying a city, we make up and return some temperature. In this case, I’ve got a hardcoded case for San Francisco.

We put all of this to work in the test, knowing that the model won’t know the current weather in a given city, and so we will have to defer to a function that we provide. It knows about the nature of the function because we’ve given it a description when we configured the FunctionCallback. It knows that the city parameter is Stringand that San Francisco is a city, so it passes the stringSan Francisco` to our function, allowing us to provide the expected response. We verify as much with the test, asserting that the response is the hardcoded magic number.

package com.example.bootiful_34.ai;

import org.junit.jupiter.api.Assertions;
import org.junit.jupiter.api.Test;
import org.springframework.ai.chat.client.ChatClient;
import org.springframework.beans.factory.annotation.Autowired;
import org.springframework.boot.test.context.SpringBootTest;

@SpringBootTest
class AiConfigurationTest {

	@Test
	void functionCallbacks(@Autowired ChatClient cc) {
		var weatherInSf = cc.prompt("give me the weather for the city of san francisco").call().content();
		Assertions.assertNotNull(weatherInSf);
		Assertions.assertTrue(weatherInSf.contains("68"));
	}

}

And just like that, we’ve given our AI model the ability to ask questions about our systems and services and to support a more agentic workflow. Easy!