Prompt 工程

Prompt设计是AI应用成功的关键，良好的Prompt设计能够显著提高AI模型的输出质量。

9.1.1 清晰明确

// 不好的Prompt
String badPrompt = "Tell me about Spring AI";

// 好的Prompt
String goodPrompt = "Explain Spring AI framework to a beginner developer, including its main features, benefits, and how to get started with a simple code example.";

9.1.2 具体详细

// 不好的Prompt
String vaguePrompt = "Write a blog post";

// 好的Prompt
String specificPrompt = "Write a 500-word blog post about Spring AI for Java developers, including code examples and best practices. The tone should be professional but accessible.";

9.1.3 上下文丰富

// 缺乏上下文的Prompt
String noContextPrompt = "What is the best way to handle errors?";

// 有上下文的Prompt
String withContextPrompt = "In a Spring AI application, what are the best practices for handling API errors from AI services? Consider retry mechanisms, fallback strategies, and user experience.";

9.1.4 角色设定

// 无角色的Prompt
String noRolePrompt = "Explain Spring Boot features";

// 有角色的Prompt
String withRolePrompt = "As a senior Spring developer with 10 years of experience, explain the key features of Spring Boot 3.x and how they improve developer productivity.";

9.2 Prompt模板

Spring AI提供了强大的Prompt模板系统，支持动态Prompt生成。

9.2.1 基础Prompt模板

@RestController
public class PromptTemplateController {

    private final ChatClient chatClient;

    public PromptTemplateController(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    @GetMapping("/prompt/template")
    public String useTemplate(@RequestParam String topic) {
        PromptTemplate promptTemplate = new PromptTemplate(
            "Tell me about {topic} in 100 words"
        );

        Prompt prompt = promptTemplate.create(Map.of("topic", topic));
        return chatClient.prompt(prompt).call().content();
    }
}

9.2.2 复杂Prompt模板

@GetMapping("/prompt/complex-template")
public String useComplexTemplate(
        @RequestParam String name,
        @RequestParam String role,
        @RequestParam String question) {

    PromptTemplate promptTemplate = new PromptTemplate(
        """
        You are {role}.
        User: {name} asks: {question}
        Please provide a detailed answer with examples.
        """
    );

    Prompt prompt = promptTemplate.create(
        Map.of("role", role, "name", name, "question", question)
    );

    return chatClient.prompt(prompt).call().content();
}

9.2.3 条件Prompt模板

@GetMapping("/prompt/conditional-template")
public String useConditionalTemplate(
        @RequestParam String language,
        @RequestParam String code) {

    PromptTemplate promptTemplate = new PromptTemplate(
        """
        Convert the following {language} code to Java:
        {code}
        """
    );

    Prompt prompt = promptTemplate.create(
        Map.of("language", language, "code", code)
    );

    return chatClient.prompt(prompt).call().content();
}

9.2.4 多语言Prompt模板

@GetMapping("/prompt/multi-language")
public Map<String, String> generateMultiLanguagePrompts(
        @RequestParam String basePrompt,
        @RequestParam List<String> languages) {

    Map<String, String> prompts = new HashMap<>();

    for (String language : languages) {
        PromptTemplate template = new PromptTemplate(
            "Translate the following to " + language + ":\n\n" + basePrompt
        );

        Prompt prompt = template.create(Map.of());
        prompts.put(language, chatClient.prompt(prompt).call().content());
    }

    return prompts;
}

9.3 Prompt优化技巧

9.3.1 分步思考（Chain of Thought）

@GetMapping("/prompt/chain-of-thought")
public String useChainOfThought(@RequestParam String complexProblem) {
    String prompt = """
        Let's solve this problem step by step:

        Problem: {problem}

        Step 1: Understand the problem
        Step 2: Identify key components
        Step 3: Develop a solution approach
        Step 4: Implement the solution
        Step 5: Verify the result

        Now, let's apply this to: {problem}
        """.replace("{problem}", complexProblem);

    return chatClient.prompt(prompt).call().content();
}

9.3.2 自我纠正

@GetMapping("/prompt/self-correction")
public String useSelfCorrection(@RequestParam String task) {
    String prompt = """
        Task: {task}

        First, provide an initial solution.
        Then, critically evaluate your solution for potential issues.
        Finally, improve the solution based on your evaluation.

        Now, let's apply this to: {task}
        """.replace("{task}", task);

    return chatClient.prompt(prompt).call().content();
}

9.3.3 示例驱动

@GetMapping("/prompt/example-driven")
public String useExampleDriven(@RequestParam String concept) {
    String prompt = """
        Explain the concept of {concept} with the following examples:

        Example 1: Simple case
        Example 2: Edge case
        Example 3: Complex case

        Now, let's apply this to: {concept}
        """.replace("{concept}", concept);

    return chatClient.prompt(prompt).call().content();
}

9.3.4 限制和约束

@GetMapping("/prompt/constraints")
public String useConstraints(@RequestParam String task) {
    String prompt = """
        Task: {task}

        Constraints:
        - Output must be under 200 words
        - Use simple language
        - Include code examples
        - Focus on practical applications

        Now, let's apply this to: {task}
        """.replace("{task}", task);

    return chatClient.prompt(prompt).call().content();
}

9.4 动态Prompt生成

9.4.1 基于上下文的Prompt生成

@Service
public class DynamicPromptService {

    private final ChatClient chatClient;

    public DynamicPromptService(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public String generateDynamicPrompt(
            String userContext,
            String task,
            Map<String, Object> constraints) {

        StringBuilder prompt = new StringBuilder();
        prompt.append("User context: ").append(userContext).append("\n\n");
        prompt.append("Task: ").append(task).append("\n\n");

        if (!constraints.isEmpty()) {
            prompt.append("Constraints:\n");
            for (Map.Entry<String, Object> entry : constraints.entrySet()) {
                prompt.append("- ").append(entry.getKey()).append(": ").append(entry.getValue()).append("\n");
            }
        }

        return chatClient.prompt(prompt.toString()).call().content();
    }
}

9.4.2 基于历史的Prompt生成

@Service
public class HistoricalPromptService {

    private final ChatClient chatClient;

    public HistoricalPromptService(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public String generateHistoricalPrompt(
            List<String> conversationHistory,
            String currentMessage) {

        StringBuilder prompt = new StringBuilder();
        prompt.append("Conversation history:\n");

        for (String message : conversationHistory) {
            prompt.append("- ").append(message).append("\n");
        }

        prompt.append("\nCurrent message: ").append(currentMessage);

        return chatClient.prompt(prompt.toString()).call().content();
    }
}

9.4.3 基于用户偏好的Prompt生成

@Service
public class PersonalizedPromptService {

    private final ChatClient chatClient;

    public PersonalizedPromptService(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public String generatePersonalizedPrompt(
            UserPreferences preferences,
            String task) {

        String prompt = "User preferences: " + preferences + "\n\n" +
                       "Task: " + task;

        return chatClient.prompt(prompt).call().content();
    }
}

public class UserPreferences {
    private String language;
    private String tone;
    private List<String> interests;
    // getters and setters
}

9.5 Prompt测试与评估

9.5.1 Prompt测试框架

public class PromptTester {

    private final ChatClient chatClient;

    public PromptTester(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public PromptTestResult testPrompt(String prompt, String expectedOutput) {
        String actualOutput = chatClient.prompt(prompt).call().content();

        double similarity = calculateSimilarity(expectedOutput, actualOutput);
        boolean passed = similarity > 0.8; // 阈值

        return new PromptTestResult(prompt, actualOutput, similarity, passed);
    }

    private double calculateSimilarity(String expected, String actual) {
        // 实现相似度计算
        return 0.9;
    }
}

public class PromptTestResult {
    private String prompt;
    private String actualOutput;
    private double similarity;
    private boolean passed;
    // getters and setters
}

9.5.2 Prompt优化循环

public class PromptOptimizer {

    private final ChatClient chatClient;
    private final PromptTester promptTester;

    public PromptOptimizer(ChatClient chatClient, PromptTester promptTester) {
        this.chatClient = chatClient;
        this.promptTester = promptTester;
    }

    public String optimizePrompt(String initialPrompt, String expectedOutput) {
        String currentPrompt = initialPrompt;
        double bestSimilarity = 0;

        for (int i = 0; i < 5; i++) { // 最多优化5次
            PromptTestResult result = promptTester.testPrompt(currentPrompt, expectedOutput);

            if (result.getSimilarity() > bestSimilarity) {
                bestSimilarity = result.getSimilarity();
                currentPrompt = refinePrompt(currentPrompt, result);
            }
        }

        return currentPrompt;
    }

    private String refinePrompt(String prompt, PromptTestResult result) {
        // 基于测试结果优化Prompt
        return prompt + " (optimized)";
    }
}

9.6 实际应用示例

9.6.1 代码生成Prompt优化

@Service
public class CodePromptOptimizer {

    private final ChatClient chatClient;

    public CodePromptOptimizer(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public String optimizeCodePrompt(
            String requirement,
            String language,
            CodeStylePreferences preferences) {

        String basePrompt = "Generate " + language + " code for: " + requirement;

        // 添加风格偏好
        String styledPrompt = basePrompt + "\nStyle: " + preferences.getStyle();

        // 添加最佳实践
        String optimizedPrompt = styledPrompt + "\nFollow best practices and include comments.";

        return optimizedPrompt;
    }
}

public class CodeStylePreferences {
    private String style;
    private boolean includeComments;
    private boolean followBestPractices;
    // getters and setters
}

9.6.2 内容生成Prompt优化

@Service
public class ContentPromptOptimizer {

    private final ChatClient chatClient;

    public ContentPromptOptimizer(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public String optimizeContentPrompt(
            String topic,
            ContentPreferences preferences) {

        String prompt = "Write content about " + topic;

        if (preferences.isIncludeExamples()) {
            prompt += " with examples";
        }

        if (preferences.isIncludeStatistics()) {
            prompt += " including relevant statistics";
        }

        if (preferences.isIncludeReferences()) {
            prompt += " and references";
        }

        return prompt;
    }
}

public class ContentPreferences {
    private boolean includeExamples;
    private boolean includeStatistics;
    private boolean includeReferences;
    // getters and setters
}

9.7 性能优化

9.7.1 Prompt缓存

@Configuration
@EnableCaching
public class PromptCacheConfig {

    @Bean
    public CacheManager cacheManager() {
        SimpleCacheManager cacheManager = new SimpleCacheManager();
        cacheManager.setCaches(Arrays.asList(
            new ConcurrentMapCache("optimized-prompts"),
            new ConcurrentMapCache("prompt-results")
        ));
        return cacheManager;
    }
}

9.7.2 Prompt预编译

@Service
public class PrecompiledPromptService {

    private final Map<String, Prompt> precompiledPrompts = new ConcurrentHashMap<>();

    public Prompt getPrecompiledPrompt(String key) {
        return precompiledPrompts.computeIfAbsent(key, this::compilePrompt);
    }

    private Prompt compilePrompt(String key) {
        // 预编译Prompt逻辑
        return new Prompt("Precompiled prompt content");
    }
}

9.8 错误处理

9.8.1 Prompt验证

@Service
public class PromptValidator {

    public boolean isValidPrompt(String prompt) {
        // 检查Prompt是否有效
        return prompt != null && !prompt.trim().isEmpty();
    }

    public List<String> validatePrompt(String prompt) {
        List<String> issues = new ArrayList<>();

        if (prompt == null || prompt.trim().isEmpty()) {
            issues.add("Prompt cannot be empty");
        }

        if (prompt.length() > 4000) {
            issues.add("Prompt is too long");
        }

        return issues;
    }
}

9.8.2 Prompt失败处理

@Service
public class PromptFallbackService {

    private final ChatClient chatClient;

    public PromptFallbackService(ChatClient chatClient) {
        this.chatClient = chatClient;
    }

    public String handlePromptFailure(String prompt) {
        try {
            return chatClient.prompt(prompt).call().content();
        } catch (Exception ex) {
            // 使用备用Prompt
            String fallbackPrompt = createFallbackPrompt(prompt);
            return chatClient.prompt(fallbackPrompt).call().content();
        }
    }

    private String createFallbackPrompt(String originalPrompt) {
        // 创建备用Prompt
        return "Simple version: " + originalPrompt;
    }
}

9.9 监控与日志

9.9.1 Prompt性能监控

@RestController
public class PromptMetricsController {

    private final MeterRegistry meterRegistry;

    public PromptMetricsController(MeterRegistry meterRegistry) {
        this.meterRegistry = meterRegistry;
    }

    @GetMapping("/prompt/metrics")
    public Map<String, Object> getMetrics() {
        return Map.of(
            "totalPrompts", meterRegistry.counter("prompt.total").count(),
            "averageProcessingTime", meterRegistry.timer("prompt.duration").mean(),
            "optimizationRate", meterRegistry.counter("prompt.optimizations").count() /
                              (meterRegistry.counter("prompt.total").count() + 1.0)
        );
    }
}

9.9.2 Prompt日志记录

@Configuration
public class PromptLogging {

    @Bean
    public ChatClient loggingChatClient(ChatClient chatClient) {
        return new ChatClient() {
            @Override
            public ChatResponse call(Prompt prompt) {
                long startTime = System.currentTimeMillis();
                try {
                    ChatResponse response = chatClient.call(prompt);
                    long duration = System.currentTimeMillis() - startTime;
                    log.info("Prompt处理 - 耗时: {}ms, Prompt: {}", duration, prompt);
                    return response;
                } catch (Exception ex) {
                    log.error("Prompt处理失败", ex);
                    throw ex;
                }
            }

            // 实现其他方法...
        };
    }
}

9.10 小结

本章详细介绍了Prompt工程的高级技术：

设计原则：掌握清晰、具体、上下文丰富、角色设定等原则
Prompt模板：学习基础和复杂Prompt模板的使用
优化技巧：了解分步思考、自我纠正、示例驱动等技巧
动态生成：掌握基于上下文、历史和用户偏好的Prompt生成
测试评估：学习Prompt测试框架和优化循环
实际应用：了解代码生成和内容生成的Prompt优化
性能优化：掌握Prompt缓存和预编译
错误处理：了解Prompt验证和失败处理
监控日志：掌握Prompt性能监控和日志记录

在下一章中，我们将学习函数调用与工具集成的技术。点击这里👇🏻获取：100万QPS短链系统、复杂的商城微服务系统、智能翻译助手AI Agent、SaaS点餐系统、刷题吧小程序、商城系统、秒杀系统、AI项目、代码生成神器、苏三demo项目、智能天气播报AI Agent、智能代码审查AI Agent等 10 个项目的：项目源代码、开发教程和技术答疑

Prompt 工程

第九章：Prompt工程

9.1 Prompt设计原则

9.1.1 清晰明确

9.1.2 具体详细

9.1.3 上下文丰富

9.1.4 角色设定

9.2 Prompt模板

9.2.1 基础Prompt模板

9.2.2 复杂Prompt模板

9.2.3 条件Prompt模板

9.2.4 多语言Prompt模板

9.3 Prompt优化技巧

9.3.1 分步思考（Chain of Thought）

9.3.2 自我纠正

9.3.3 示例驱动

9.3.4 限制和约束

9.4 动态Prompt生成

9.4.1 基于上下文的Prompt生成

9.4.2 基于历史的Prompt生成

9.4.3 基于用户偏好的Prompt生成

9.5 Prompt测试与评估

9.5.1 Prompt测试框架

9.5.2 Prompt优化循环

9.6 实际应用示例

9.6.1 代码生成Prompt优化

9.6.2 内容生成Prompt优化

9.7 性能优化

9.7.1 Prompt缓存

9.7.2 Prompt预编译

9.8 错误处理

9.8.1 Prompt验证

9.8.2 Prompt失败处理

9.9 监控与日志

9.9.1 Prompt性能监控

9.9.2 Prompt日志记录

9.10 小结