注目イベント！
アドベントカレンダー2024開催中！
一年を締めくくる特別なイベント、アドベントカレンダーを今年も開催しています！
初心者からベテランまで楽しめる内容で、毎日新しい技術トピックをお届けします。
詳細はこちらから！
event banner

Trying Out LangChain4j, the Java Library for LangChain

日本語に切り替え
| 15 min read
Author: kotaro-miura kotaro-miuraの画像
#RAG #OpenAI #GPT #LangChain
Information

To reach a broader audience, this article has been translated from Japanese.
You can find the original version here.

Introduction

#

Hello. This is Miura from the Digital Strategy Support Division.

This time, I would like to introduce LangChain4j, the Java library for LangChain, which I have tried out.

LangChain is a framework that consolidates useful features for application development using LLM (Large Language Models). It supports many features such as utilizing various language models with a unified interface, prompt templates, conversation history storage, embeddings, and connections to vector DBs. LangChain is provided as libraries for various programming languages, with Python and JavaScript/TypeScript being particularly popular. However, since I often use Java personally, I decided to try out the LangChain library for Java this time.

There seem to be several open-source LangChain libraries for Java, but I would like to use LangChain4j, which is the most popular in the Maven repository.

Package

#

Import the following package.

<dependency>
  <groupId>dev.langchain4j</groupId>
  <artifactId>langchain4j</artifactId>
  <version>0.30.0</version>
</dependency>

If you are using OpenAI's API for questions, you will need the following package.

<dependency>
  <groupId>dev.langchain4j</groupId>
  <artifactId>langchain4j-open-ai</artifactId>
  <version>0.30.0</version>
</dependency>
Information

If you are using other generative AI services such as Azure OpenAI, Amazon Bedrock, or Anthropic, you need to import the necessary packages for each. You can check the list of supported services at the following link.
Supported LLM Integrations

Asking the Chat Model

#

The simplest code to ask the chat model is as follows.

OpenAiChatModel model = OpenAiChatModel.withApiKey("API_KEY");
String answer = model.generate("こんにちは");
System.out.println(answer);
// こんにちは！何かお手伝いできることはありますか？

By default, it asks the gpt-3.5-turbo model.

To specify the model and other parameters, use the builder method as follows.

OpenAiChatModel model = OpenAiChatModel.builder()
        .apiKey("API_KEY")
        .modelName(OpenAiChatModelName.GPT_3_5_TURBO)
        .build();
String answer = model.generate("こんにちは");

You can also ask questions including past responses as follows.

OpenAiChatModel model = OpenAiChatModel.builder()
        .apiKey("API_KEY")
        .modelName("gpt-4")
        .build();
var userMsg1 = UserMessage.from("こんにちは、私の名前は豆田豆蔵です");
AiMessage aiMsg1 = model.generate(userMsg1).content();
System.out.println(aiMsg1.text());
// こんにちは、豆田豆蔵さん。何かお手伝いできることがありますか？

var userMsg2 = UserMessage.from("私の名前はなんですか？");
AiMessage aiMsg2 = model.generate(userMsg1, aiMsg1, userMsg2).content();
System.out.println(aiMsg2.text());
// あなたの名前は豆田豆蔵さんです。

Stream Output

#

To stream the response content token by token, the code is as follows.

StreamingChatLanguageModel model = OpenAiStreamingChatModel.withApiKey("API_KEY");

String userMessage = "こんにちは";

model.generate(userMessage, new StreamingResponseHandler<AiMessage>() {

    @Override
    public void onNext(String token) {
        System.out.println("onNext: " + token);
    }

    @Override
    public void onComplete(Response<AiMessage> response) {
        System.out.println("onComplete: " + response);
    }

    @Override
    public void onError(Throwable error) {
        error.printStackTrace();
    }
});

/* Output result
onNext: 
onNext: こんにちは
onNext: ！
onNext: お
onNext: 元
onNext: 気
onNext: です
onNext: か
onNext: ？
onNext: お
onNext: 手
onNext: 伝
onNext: い
onNext: で
onNext: き
onNext: る
onNext: こ
onNext: と
onNext: が
onNext: あ
onNext: れ
onNext: ば
onNext: お
onNext: 知
onNext: ら
onNext: せ
onNext: ください
onNext: 。
onComplete: Response { content = AiMessage { text = "こんにちは！お元気ですか？お手伝いできることがあればお知らせください。" toolExecutionRequests = null }, tokenUsage = TokenUsage { inputTokenCount = 8, outputTokenCount = 28, totalTokenCount = 36 }, finishReason = STOP }
*/

You need to pass an implementation instance of StreamingResponseHandler to the generate method. Implement the processing for each token in onNext, the processing when output is complete in onComplete, and the processing when an error occurs in onError.

AI Service

#

LangChain4j provides a module called AI Service to make it easier to implement features using generative AI. AI Service supports features such as conversation history retention, Tools, and RAG.

Question

#

The following code creates an AI Service Assistant that simply asks questions.

interface Assistant {
    String chat(String userMessage);
}

OpenAiChatModel model = OpenAiChatModel.withApiKey("API_KEY");
Assistant assistant = AiServices.create(Assistant.class, model);
String answer = assistant.chat("こんにちは");

Prepare an interface called Assistant and let it have a chat method with String as input and output.

This Assistant can ask questions by creating an instance using AiServices#create.

LangChain4j automatically determines the processing content of the Assistant#chat method, and the implementation content changes automatically depending on the input and output type patterns of this method. This time, since the method has both input and output as String, it is determined to perform simple question processing.

System Message

#

By adding the @SystemMessage annotation, it will respond according to the instructions in the system message.

The following creates an AI Service Friend that gives casual responses to questions.

interface Friend {
    @SystemMessage("You are my best friend. Please give casual responses.")
    String chat(String userMessage);
}

OpenAiChatModel model = OpenAiChatModel.withApiKey("API_KEY");
Friend friend = AiServices.create(Friend.class, model);
String answer = assistant.chat("こんにちは");
// おーい、どうも！何か面白いことあった？

Parsing AI Output

#

You can convert the content of AI responses into Java objects.

For example, the following code creates an AI Service SentimentAnalyzer that outputs the classification results of the emotions contained in a sentence as Enum or Boolean.

enum Sentiment {
    POSITIVE, NEUTRAL, NEGATIVE
}

interface SentimentAnalyzer {
    @UserMessage("Please analyze the sentiment of the following sentence: {{it}}")
    Sentiment analyzeSentimentOf(String text);

    @UserMessage("Does the following sentence express positive sentiment?: {{it}}")
    boolean isPositive(String text);
}

OpenAiChatModel model = OpenAiChatModel.withApiKey("API_KEY");
SentimentAnalyzer sentimentAnalyzer = AiServices.create(SentimentAnalyzer.class, model);

Sentiment sentiment = sentimentAnalyzer.analyzeSentimentOf("これは素晴らしい！");
System.out.println(sentiment);
// POSITIVE

boolean isPositive = sentimentAnalyzer.isPositive("怖い");
System.out.println(isPositive);
// false

It can also parse into custom POJOs.

The following creates an AI Service PersonExtractor that parses the person information contained in a sentence into the Person class.

record Person {
    String firstName;
    String lastName;
    LocalDate birthDate;
    Address address;
}

record Address {
    String street;
    Integer streetNumber;
    String city;
}

interface PersonExtractor {
    @UserMessage("Please extract the person information from the following sentence: {{it}}")
    Person extractPersonFrom(String text);
}

OpenAiChatModel model = OpenAiChatModel.builder()
        .apiKey("API_KEY")
        .modelName(OpenAiChatModelName.GPT_3_5_TURBO)
        .responseFormat("json_object")
        .build();
PersonExtractor personExtractor = AiServices.create(PersonExtractor.class, model);

String text = """
        こんにちは私の名前は豆田豆蔵です。2024年5月1日生まれです。
        東京都新宿区西新宿二丁目1番1号 新宿三井ビルディング34階に住んでいます。
        """;

Person person = personExtractor.extractPersonFrom(text);
System.out.println(person.toString());
// Person[firstName=豆蔵, lastName=豆田, birthDate=2024-05-01, address=Address[street=西新宿二丁目, streetNumber=1, city=東京都新宿区]]

Memory

#

You can ask questions including past responses using ChatMemory.

You can set it to consider the last N messages or the last N tokens of messages.

The following uses MessageWindowChatMemory to consider the last 10 messages when answering.

ChatMemory chatMemory = MessageWindowChatMemory.withMaxMessages(10);

Assistant assistant = AiServices.builder(Assistant.class)
        .chatLanguageModel(OpenAiChatModel.withApiKey("API_KEY"))
        .chatMemory(chatMemory)
        .build();

String answer = assistant.chat("こんにちは、私の名前は豆田豆蔵です");
System.out.println(answer);
// 初めまして、豆田豆蔵さん。どのようなお仕事をされていますか？

String answerWithName = assistant.chat("私の名前がわかりますか？");
System.out.println(answerWithName);
// はい、お名前は豆田豆蔵さんですね。どのようなお話でもお聞きしますので、お気軽にどうぞ。

Tools (Function Calling)

#

Function Calling can be executed by preparing a class with methods annotated with @Tool.

If the response content of the question prompts the execution of a Tool, it will automatically execute and relay the results.

The following code creates an AI Service MathGenius that performs addition and square root calculations.

interface MathGenius {
    String ask(String question);
}

class Calculator {
    @Tool
    public double add(int a, int b) {
        return a + b;
    }

    @Tool
    public double squareRoot(double x) {
        return Math.sqrt(x);
    }
}

OpenAiChatModel model = OpenAiChatModel.withApiKey("API_KEY");
MathGenius mathGenius = AiServices.builder(MathGenius.class)
        .chatLanguageModel(model)
        .tools(new Calculator())
        .build();

String answer = mathGenius.ask("What is the square root of 475695037565?");
System.out.println(answer);
// The square root of 475695037565 is approximately 689706.49.

RAG

#

RAG is often used as a method to generate responses based on proprietary data. Although the explanation of RAG itself is omitted in this article, LangChain4j provides interfaces to simplify the implementation of RAG, so I will introduce them.

The following code implements the basic flow of RAG.

OpenAiChatModel model = OpenAiChatModel.withApiKey("API_KEY");

DocumentParser documentParser = new ApacheTikaDocumentParser();
Document document = FileSystemDocumentLoader.loadDocument("DOCUMENT_PATH", documentParser);

DocumentSplitter splitter = DocumentSplitters.recursive(300, 0);

EmbeddingModel embeddingModel = new BgeSmallEnV15QuantizedEmbeddingModel();

EmbeddingStore<TextSegment> embeddingStore = new InMemoryEmbeddingStore<>();

EmbeddingStoreIngestor ingestor = EmbeddingStoreIngestor.builder()
        .documentSplitter(splitter)
        .embeddingModel(embeddingModel)
        .embeddingStore(embeddingStore)
        .build();

ingestor.ingest(document);

ContentRetriever contentRetriever = EmbeddingStoreContentRetriever.builder()
        .embeddingStore(embeddingStore)
        .embeddingModel(embeddingModel)
        .maxResults(2)
        .minScore(0.5)
        .build();

ChatMemory chatMemory = MessageWindowChatMemory.withMaxMessages(10);

var assistant = AiServices.builder(Assistant.class)
        .chatLanguageModel(model)
        .contentRetriever(contentRetriever)
        .chatMemory(chatMemory)
        .build();

String answer = assistant.chat("Question text");
System.out.println(answer);

DocumentParser is an interface for extracting text information from files. It supports not only text files but also pdfs, PowerPoints, etc.

DocumentLoader is an interface for loading files. It also supports loading from AWS S3 and Azure blob.

DocumentSplitter is an interface for splitting text into chunks. You can set how many characters to split and how many characters to overlap between chunks.

EmbeddingModel is an interface that represents a model that converts text into vectors. It can use models that process locally or public models such as OpenAI.

EmbeddingStore is an interface that represents a data source that stores vector data. It can be stored in memory, and it also supports many DBs such as PostgreSQL and Redis.

By using EmbeddingStoreIngestor, it splits the document into text chunks, converts them into vectors, and stores them in the vector store.

Specify the embedding model that converts the question text into vectors and the vector store to search in ContentRetriever.

Although there are quite a few necessary modules, I think it makes it possible to avoid implementing most of the processes required for RAG manually.

Advanced RAG

#

Various methods have been devised to improve the accuracy of responses by breaking down the RAG process and adding ingenuity to each process.

LangChain4j provides interfaces for each decomposed process, making it easier to implement methods to improve RAG accuracy.

The conceptual diagram on the next page may help you get an image of the decomposition of the RAG process.

LangChain4j Advanced RAG

  • QueryTransformer
    • Transforming search queries for the vector store
    • Example usage: Query Compression (rephrasing past conversation history into a single question suitable for the search query)
  • QueryRouter
    • Selecting which vector store to search when there are multiple vector stores
    • Example usage: Query Routing, skipping the search process
  • ContentAggregator
    • Aggregating the extracted data
    • Example usage: Re-rank (reordering the extracted chunks in order of relevance to the question text)
  • ContentInjector
    • Adding information to the extracted data
    • Example usage: Adding metadata (document name, split index)

Implementation examples are also provided, which helps in understanding.

Conclusion

#

My impression after trying it out is that it is very helpful to be able to easily implement Function Calling and RAG.

This time, I used OpenAI's public API for questions, but I would like to try other services such as Azure OpenAI and Amazon Bedrock.

Looking at the commits on LangChain4j's GitHub, it seems that development is still actively ongoing, so I would like to keep an eye on future developments.

豆蔵では共に高め合う仲間を募集しています！

recruit

具体的な採用情報はこちらからご覧いただけます。