agents-flex/docs/core/llms.md

Large Language Models (LLMs)

Agents-Flex provides an abstract implementation interface for Large Language Models (LLMs) called Llm.java, supporting two different dialogue modes: chat and chatStream.

For various vendors, Agents-Flex offers different implementation classes and communication protocols, including HTTP, SSE, and WebSocket clients.

Chat dialogue

In AI large language model chat conversations, we need to consider several different scenarios:

• Simple conversation
• Historical conversation
• Function Calling

These capabilities are determined by prompts, so Agents-Flex provides three implementations of prompts:

• SimplePrompt: Used for simple conversation scenarios
• HistoriesPrompt: Used for historical conversation scenarios
• FunctionPrompt: Used for Function Calling scenarios

During the interaction between prompts and large models, messages are exchanged. Therefore, Agents-Flex also provides different message implementations:

• AiMessage: The message returned by the large model, which includes not only the message content but also data such as token consumption.
• FunctionMessage: A subclass of AiMessage, returned when using FunctionPrompt in the chat method.
• HumanMessage: Represents messages input by users during conversations.
• SystemMessage: Represents system messages, often used to inform the large language model's role, for fine-tuning prompts.

Example Code

Simple Dialogue

public static void main(String[] args) {
    Llm llm = new OpenAiLlm.of("sk-rts5NF6n*******");

    Prompt prompt = new SimplePrompt("what is your name?");
    String response = llm.chat(prompt);

    System.out.println(response);
}

Historical Dialogue

public static void main(String[] args) {
    Llm llm = new OpenAiLlm.of("sk-rts5NF6n*******");

    HistoriesPrompt prompt = new HistoriesPrompt();
    prompt.addMessage(new SystemMessage("You are now a database development engineer...."));
    prompt.addMessage(new HumanMessage("Please provide the DDL content for...."));

    String response = llm.chat(prompt);

    System.out.println(response);
}

Function Calling

Utility class definition：

public class WeatherUtil {

    @FunctionDef(name = "get_the_weather_info", description = "get the weather info")
    public static String getWeatherInfo(
        @FunctionParam(name = "city", description = "the city name") String name) {
        //Here, we should retrieve API information through third-party interfaces.
        return name + "The weather is cloudy turning to overcast. ";
    }
}

Create FunctionPrompt and pass it to the large model through the chat method:

public static void main(String[] args) {
    OpenAiLlmConfig config = new OpenAiLlmConfig();
    config.setApiKey("sk-rts5NF6n*******");

    OpenAiLlm llm = new OpenAiLlm(config);

    FunctionPrompt prompt = new FunctionPrompt("Today, how's the weather in New York?", WeatherUtil.class);
    FunctionResultResponse response = llm.chat(prompt);

    //Execute utility class method to obtain result.
    Object result = response.invoke();

    System.out.println(result);
    //"The weather in New York is cloudy turning overcast."
}