ChatGPT Spring Boot Starter =========================== Spring Boot ChatGPT starter with ChatGPT chat and functions support. # Features * Base on Spring Boot 3.0+ * Async with Spring Webflux * Support ChatGPT Chat Stream * Support ChatGPT functions: `@GPTFunction` annotation * Support structured output: `@StructuredOutput` annotation for record * Prompt Management: load prompt templates from `prompt.properties` with `@PropertyKey`, and friendly with IntelliJ IDEA * Prompt as Lambda: convert prompt template to lambda expression and call it with FP style * ChatGPT interface: Declare ChatGPT service interface with `@ChatGPTExchange` and `@ChatCompletion` annotations. * No third-party library: base on Spring 6 HTTP interface * GraalVM native image support * Azure OpenAI support # Get Started ### Add dependency Add `chatgpt-spring-boot-starter` dependency in your pom.xml. ```xml org.mvnsearch chatgpt-spring-boot-starter 0.8.0 ``` ### Adjust configuration Add `openai.api.key` in `application.properties`: ```properties # OpenAI API Token, or you can set environment variable OPENAI_API_KEY openai.api.key=sk-proj-xxxx ``` If you want to use Azure OpenAI, you can add `openai.api.url` in `application.properties`: ```properties openai.api.key=1138xxxx9037 openai.api.url=https://YOUR_RESOURCE_NAME.openai.azure.com/openai/deployments/YOUR_DEPLOYMENT_NAME/chat/completions?api-version=2023-05-15 ``` ### Call ChatGPT Service ```java @RestController public class ChatRobotController { @Autowired private ChatGPTService chatGPTService; @PostMapping("/chat") public Mono chat(@RequestBody String content) { return chatGPTService.chat(ChatCompletionRequest.of(content)) .map(ChatCompletionResponse::getReplyText); } @GetMapping("/stream-chat") public Flux streamChat(@RequestParam String content) { return chatGPTService.stream(ChatCompletionRequest.of(content)) .map(ChatCompletionResponse::getReplyText); } } ``` # ChatGPT Service Interface ChatGPT service interface is almost like [Spring 6 HTTP Interface](https://docs.spring.io/spring-framework/reference/integration/rest-clients.html#rest-http-interface). You can declare a ChatGPT service interface with `@ChatGPTExchange` annotation, and declare completion methods with `@ChatCompletion` annotation, then you just call service interface directly. ```java @GPTExchange public interface GPTHelloService { @ChatCompletion("You are a language translator, please translate the below text to Chinese.\n") Mono translateIntoChinese(String text); @ChatCompletion("You are a language translator, please translate the below text from {0} to {1}.\n {2}") Mono translate(String sourceLanguage, String targetLanguage, String text); @Completion("please complete poem: {0}") Mono completePoem(String text); } ``` Create ChatGPT interface service bean: ``` @Bean public GPTHelloService gptHelloService(ChatGPTServiceProxyFactory proxyFactory) { return proxyFactory.createClient(GPTHelloService.class); } ``` # ChatGPT functions * Create a Spring Bean with `@Component`. Annotate GPT functions with `@GPTFunction` annotation, and annotate function parameters with `@Parameter` annotation. `@Nonnull` means that the parameter is required. ```java import jakarta.annotation.Nonnull; @Component public class GPTFunctions { public record SendEmailRequest( @Nonnull @Parameter("Recipients of email") List recipients, @Nonnull @Parameter("Subject of email") String subject, @Parameter("Content of email") String content) { } @GPTFunction(name = "send_email", value = "Send email to receiver") public String sendEmail(SendEmailRequest request) { System.out.println("Recipients: " + String.join(",", request.recipients)); System.out.println("Subject: " + request.subject); System.out.println("Content:\n" + request.content); return "Email sent to " + String.join(",", request.recipients) + " successfully!"; } public record SQLQueryRequest( @Parameter(required = true, value = "SQL to query") String sql) { } @GPTFunction(name = "execute_sql_query", value = "Execute SQL query and return the result set") public String executeSQLQuery(SQLQueryRequest request) { System.out.println("Execute SQL: " + request.sql); return "id, name, salary\n1,Jackie,8000\n2,Libing,78000\n3,Sam,7500"; } } ``` * Call GPT function by `response.getReplyCombinedText()` or `chatMessage.getFunctionCall().getFunctionStub().call()`: ```java public class ChatGPTServiceImplTest { @Test public void testChatWithFunctions() throws Exception { final String prompt = "Hi Jackie, could you write an email to Libing(libing.chen@gmail.com) and Sam(linux_china@hotmail.com) and invite them to join Mike's birthday party at 4 pm tomorrow? Thanks!"; final ChatCompletionRequest request = ChatCompletionRequest.functions(prompt, List.of("send_email")); final ChatCompletionResponse response = chatGPTService.chat(request).block(); // display reply combined text with function call System.out.println(response.getReplyCombinedText()); // call function manually for (ChatMessage chatMessage : response.getReply()) { final FunctionCall functionCall = chatMessage.getFunctionCall(); if (functionCall != null) { final Object result = functionCall.getFunctionStub().call(); System.out.println(result); } } } @Test public void testExecuteSQLQuery() { String context = "You are SQL developer. Write SQL according to requirements, and execute it in MySQL database."; final String prompt = "Query all employees whose salary is greater than the average."; final ChatCompletionRequest request = ChatCompletionRequest.functions(prompt, List.of("execute_sql_query")); // add prompt context as system message request.addMessage(ChatMessage.systemMessage(context)); final ChatCompletionResponse response = chatGPTService.chat(request).block(); System.out.println(response.getReplyCombinedText()); } } ``` **Note**: `@GPTExchange` and `@ChatCompletion` has functions built-in, so you just need to fill functions parameters. ### ChatGPT Functions use cases: * Structure Output: such as SQL, JSON, CSV, YAML etc., then delegate functions to process them. * Commands: such as send_email, post on Twitter. * DevOps: such as generate K8S yaml file, then call K8S functions to deploy it. * Search Matching: bind search with functions, such as search for a book, then call function to show it. * Spam detection: email spam, advertisement spam etc * PipeLine: you can think function as a node in pipeline. After process by function, and you can pass it to ChatGPT again. * Data types supported: `string`, `number`, `integer`, `array`. Nested `object` not supported now! If you want to have a simple test for ChatGPT functions, you can install [ChatGPT with Markdown JetBrains IDE Plugin](https://plugins.jetbrains.com/plugin/21671-chatgpt-with-markdown), and take a look at [chat.gpt file](./chat.gpt). # Structured Output Please refer [OpenAI Structured Outputs](https://platform.openai.com/docs/guides/structured-outputs) for detail. First you need to define record for structured output: ```java @StructuredOutput(name = "java_example") public record JavaExample(@Nonnull @Parameter("explanation") String explanation, @Nonnull @Parameter("answer") String answer, @Nonnull @Parameter("code") String code, @Nonnull @Parameter("dependencies") List dependencies) { } ``` Then you can use structured output record as return type as following: ```java @ChatCompletion(system = "You are a helpful Java language assistant.") Mono generateJavaExample(String question); @ChatCompletion(system = "You are a helpful assistant.", user = "Say hello to {0}!") Mono hello(String word); ``` **Attention**: if the return type is not `Mono`, and it means structured output. # Prompt Templates How to manage prompts in Java? Now ChatGPT starter adopts `prompts.properties` to save prompt templates, and uses MessageFormat to format template value.`PromptPropertiesStoreImpl` will load all ` prompts.properties` files from classpath. You can extend `PromptStore` to load prompts from database or other sources. You can load prompt template by [PromptManager](src/main/java/org/mvnsearch/chatgpt/spring/service/PromptManager.java). Tips: * Prompt template code completion: support by `@PropertyKey(resourceBundle = PROMPTS_FQN)` * `@ChatCompletion` annotation has built-in prompt template support for `user`,`system` and `assistant` messages. * Prompt value could be from classpath and URL: `conversation=classpath:///conversation-prompt.txt` or `conversation=https://example.com/conversation-prompt.txt` ### Prompt Template as Lambda For some case you want to use prompt template as lambda, such as translate first, then send it as email. You can declare prompt as function and chain them together. ```java public class PromptLambdaTest { @Test public void testPromptAsFunction() { Function> translateIntoChineseFunction = chatGPTService.promptAsLambda("translate-into-chinese"); Function> sendEmailFunction = chatGPTService.promptAsLambda("send-email"); String result = Mono.just("Hi Jackie, could you write an email to Libing(libing.chen@exaple.com) and Sam(linux_china@example.com) and invite them to join Mike's birthday party at 4 pm tomorrow? Thanks!") .flatMap(translateIntoChineseFunction) .flatMap(sendEmailFunction) .block(); System.out.println(result); } } ``` To keep placeholders safe in prompt template, you can use record as Lambda parameter. ```java public class PromptTest { public record TranslateRequest(String from, String to, String text) { } @Test public void testLambdaWithRecord() { Function> translateFunction = chatGPTService.promptAsLambda("translate"); String result = Mono.just(new TranslateRequest("Chinese", "English", "你好！")) .flatMap(translateFunction) .block(); System.out.println(result); } } ``` # [Batch API](https://platform.openai.com/docs/guides/batch) - Convert multi requests to JSONL format - Upload JSONL file to OpenAI - Create batch with file id ``` @Autowired private OpenAIFileAPI openAIFileAPI; @Autowired private OpenAIBatchAPI openAIBatchAPI; @Test public void testUpload() { String jsonl = Stream.of("What's Java Language?", "What's Kotlin Language?") .map(ChatCompletionRequest::of) .map(ChatCompletionBatchRequest::new) .map(this::toJson) .filter(Strings::isNotBlank) .collect(Collectors.joining("\n")); Resource resource = new ByteArrayResource(jsonl.getBytes()); FileObject fileObject = openAIFileAPI.upload("batch", resource).block(); CreateBatchRequest createBatchRequest = new CreateBatchRequest(fileObject.getId()); BatchObject batchObject = openAIBatchAPI.create(createBatchRequest).block(); } ``` After `completion_window(24h)`, and you can call `openAIBatchAPI.retrieve(batchId)` to get the `BatchObject`. Get `BatchObject.outputFileId` and call `OpenAIFileAPI.retrieve(outputFileId)` to get jsonl response, and use follow code to parse every chat response. ``` List lines = new BufferedReader(new InputStreamReader(inputStream)).lines().toList(); for (String line : lines) { if (line.startsWith("{")) { ChatCompletionBatchResponse response = objectMapper.readValue(line, ChatCompletionBatchResponse.class); System.out.println(response.getCustomId()); } } ``` # FAQ ### How to integrate with DeepSeek? ```properties openai.api.url=https://api.deepseek.com/v1 openai.api.key=sk-xxxx openai.model=deepseek-chat ``` ### OpenAI REST API proxy Please refer [OpenAIProxyController](src/test/java/org/mvnsearch/chatgpt/demo/OpenAIProxyController.java). ```java @RestController public class OpenAIProxyController { @Autowired private OpenAIChatAPI openAIChatAPI; @PostMapping("/v1/chat/completions") public Publisher completions(@RequestBody ChatCompletionRequest request) { return openAIChatAPI.proxy(request); } } ``` Of course, you can use standard URL `http://localhost:8080/v1/chat/completions` to call Azure OpenAI API. ### How to use ChatGPT with Spring Web? Now ChatGPT starter use Reactive style API, and you know Reactive still hard to understand. Could ChatGPT starter work with Spring Web? Yes, you can use `Mono` or `Flux` with Spring Web and Virtual Threads, please refer [Support for Virtual Threads on Spring Boot 3.2](https://github.com/spring-projects/spring-boot/wiki/Spring-Boot-3.2-Release-Notes#support-for-virtual-threads) for details. # Building the Code The code uses the Spring Java Formatter Maven plugin, which keeps the code consistent. In order to build the code, run: ```shell {#javaformat} ./mvnw spring-javaformat:apply ``` This will ensure that all contributions have the exact same code formatting, allowing us to focus on bigger issues, like functionality, # References * [OpenAI chat API](https://platform.openai.com/docs/api-reference/chat) * [Spring Boot Docs](https://docs.spring.io/spring-boot/docs/current/reference/html/) * [Spring Boot Webflux](https://docs.spring.io/spring-framework/reference/web/webflux.html) * [Spring 6 HTTP interface](https://docs.spring.io/spring-framework/reference/integration/rest-clients.html#rest-http-interface) * [Properties File Format](https://docs.oracle.com/cd/E23095_01/Platform.93/ATGProgGuide/html/s0204propertiesfileformat01.html) * [MessageFormat JavaDoc](https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/text/MessageFormat.html) * [OpenAI Prompt engineering](https://platform.openai.com/docs/guides/prompt-engineering)