本文介绍: First, generates a ChatParameter instance that represents the given class. Then, invoke the OpenAI API and return the function name and parameters. Finally, execute functions through the FunctionExecutor.

From the project: https://github.com/HamaWhiteGG/langchain-java

Please refer to the ChatFunctionTest for the complete test code.

1. Core main process

2. Stepby-step introduction

2.1 Declare function parameters and response

public record Weather(
        @JsonProperty(required = true)
        @JsonPropertyDescription("The city and state, e.g. San Francisco, CA")
        String location,

        @JsonPropertyDescription("The temperature unit")
        WeatherUnit unit
) {}


public enum WeatherUnit {

    CELSIUS,
    FAHRENHEIT
}


@Data
@Builder
public class WeatherResponse {

    public String location;

    public WeatherUnit unit;

    public int temperature;

    public String description;
}


WeatherResponse getCurrentWeather(Weather weather) {
    // mock function
    return WeatherResponse.builder()
            .location(weather.location())
            .unit(weather.unit())
            .temperature(new Random().nextInt(50))
            .description("sunny")
            .build();
}

2.2 Generates a ChatParameter instance that represents the given class

Automatically convert Weather.class to ChatParameter using the ChatParameterUtils.generate method.

ChatFunction.ChatParameter chatParameter = ChatParameterUtils.generate(Weather.class);

The output of converting chatParameter to JSON String is as follows:

{
    "type" : "object",
    "properties" : {
        "location" : {
            "type" : "string",
            "description" : "The city and state, e.g. San Francisco, CA"
        },
        "unit" : {
            "type" : "string",
            "description" : "The temperature unit",
            "enum" : [
                "celsius",
                "fahrenheit"
            ]
        }
    },
    "required" : [
        "location"
    ]
}

2.3 Invoke the OpenAI API and return the function name and parameters.

ChatFunction chatFunction = ChatFunction.builder()
        .name(functionName)
        .description("Get the current weather in a given location")
        .parameters(ChatParameterUtils.generate(Weather.class))
        .build();

Message message = Message.of("What is the weather like in Boston?");

ChatCompletion chatCompletion = ChatCompletion.builder()
        .model("gpt-4")
        .temperature(0)
        .messages(List.of(message))
        .tools(List.of(new Tool(chatFunction)))
        .toolChoice("auto")
        .build();

ChatCompletionResp response = client.createChatCompletion(chatCompletion);
ChatChoice chatChoice = response.getChoices().get(0);

The output of converting chatChoice to JSON String is as follows:

{
    "index":0,
    "message":{
        "role":"assistant",
        "content":null,
        "tool_calls":[
            {
                "id":"call_aS4LspVVBkE8uIiBSLZWXe6O",
                "type":"function",
                "function":{
                    "name":"get_current_weather",
                    "arguments":"{n  "location": "Boston, MA"n}"
                }
            }
        ]
    },
    "finish_reason":"tool_calls"
}

2.4 Execute functions through the FunctionExecutor.

The code for FunctionExecutor is as follows, allowing the caller to avoid explicit type casting through generics, making the code more concise and elegant.

public class FunctionExecutor {

    private static final Map<String, Function<?, ?>> FUNCTION_MAP = new HashMap<>(16);

    public static <T> void register(String functionName, Function<T, ?> function) {
        FUNCTION_MAP.put(functionName, function);
    }

    @SuppressWarnings("unchecked")
    public static <T, R> R execute(String functionName, Class<T> clazz, String arguments) throws ClassCastException {
        Function<T, R> function = (Function<T, R>) FUNCTION_MAP.get(functionName);
        if (function == null) {
            throw new IllegalArgumentException("No function registered with name: " + functionName);
        }
        T input = convertFromJsonStr(arguments, clazz);
        return function.apply(input);
    }
}

The caller’s code is as follows:

Function function = chatChoice.getMessage().getToolCalls().get(0).getFunction();

// execute function
WeatherResponse weatherResponse = FunctionExecutor.execute(function.getName(), Weather.class, function.getArguments());
System.out.println(weatherResponse)

The final result is:

WeatherResponse(location=Boston, MA, unit=null, temperature=2, description=sunny)

发表回复

您的邮箱地址不会被公开。 必填项已用 * 标注