Introduction: OpenAI GPT-3 Api Client in Java
More: Author   ReportBugs   

Maven Central

⚠️OpenAI has deprecated all Engine-based APIs. See Deprecated Endpoints below for more info.

Java libraries for using OpenAI's GPT apis. Supports GPT-3, ChatGPT, and GPT-4.

Includes the following artifacts:

  • api : request/response POJOs for the GPT APIs.
  • client : a basic retrofit client for the GPT endpoints, includes the api module
  • service : A basic service class that creates and calls the client. This is the easiest way to get started.

as well as an example project using the service.

Supported APIs

Deprecated by OpenAI



implementation 'com.theokanning.openai-gpt3-java:<api|client|service>:<version>'




Data classes only

If you want to make your own client, just import the POJOs from the api module. Your client will need to use snake case to work with the OpenAI API.

Retrofit client

If you're using retrofit, you can import the client module and use the OpenAiApi.
You'll have to add your auth token as a header (see AuthenticationInterceptor) and set your converter factory to use snake case and only include non-null fields.


If you're looking for the fastest solution, import the service module and use OpenAiService.

⚠️The OpenAiService in the client module is deprecated, please switch to the new version in the service module.

OpenAiService service = new OpenAiService("your_token");
CompletionRequest completionRequest = CompletionRequest.builder()
        .prompt("Somebody once told me the world is gonna roll me")

Customizing OpenAiService

If you need to customize OpenAiService, create your own Retrofit client and pass it in to the constructor. For example, do the following to add request logging (after adding the logging gradle dependency):

ObjectMapper mapper = defaultObjectMapper();
OkHttpClient client = defaultClient(token, timeout)
Retrofit retrofit = defaultRetrofit(client, mapper);

OpenAiApi api = retrofit.create(OpenAiApi.class);
OpenAiService service = new OpenAiService(api);

Adding a Proxy

To use a proxy, modify the OkHttp client as shown below:

ObjectMapper mapper = defaultObjectMapper();
Proxy proxy = new Proxy(Proxy.Type.HTTP, new InetSocketAddress(host, port));
OkHttpClient client = defaultClient(token, timeout)
Retrofit retrofit = defaultRetrofit(client, mapper);
OpenAiApi api = retrofit.create(OpenAiApi.class);
OpenAiService service = new OpenAiService(api);


You can create your functions and define their executors easily using the ChatFunction class, along with any of your custom classes that will serve to define their available parameters. You can also process the functions with ease, with the help of an executor called FunctionExecutor.

First we declare our function parameters:

public class Weather {
    @JsonPropertyDescription("City and state, for example: León, Guanajuato")
    public String location;
    @JsonPropertyDescription("The temperature unit, can be 'celsius' or 'fahrenheit'")
    @JsonProperty(required = true)
    public WeatherUnit unit;
public enum WeatherUnit {
public static class WeatherResponse {
    public String location;
    public WeatherUnit unit;
    public int temperature;
    public String description;

    // constructor

Next, we declare the function itself and associate it with an executor, in this example we will fake a response from some API:

        .description("Get the current weather of a location")
        .executor(Weather.class, w -> new WeatherResponse(w.location, w.unit, new Random().nextInt(50), "sunny"))

Then, we employ the FunctionExecutor object from the 'service' module to assist with execution and transformation into an object that is ready for the conversation:

List<ChatFunction> functionList = // list with functions
FunctionExecutor functionExecutor = new FunctionExecutor(functionList);

List<ChatMessage> messages = new ArrayList<>();
ChatMessage userMessage = new ChatMessage(ChatMessageRole.USER.value(), "Tell me the weather in Barcelona.");
ChatCompletionRequest chatCompletionRequest = ChatCompletionRequest
        .functionCall(new ChatCompletionRequestFunctionCall("auto"))

ChatMessage responseMessage = service.createChatCompletion(chatCompletionRequest).getChoices().get(0).getMessage();
ChatFunctionCall functionCall = responseMessage.getFunctionCall(); // might be null, but in this case it is certainly a call to our 'get_weather' function.

ChatMessage functionResponseMessage = functionExecutor.executeAndConvertToMessageHandlingExceptions(functionCall);

Note: The FunctionExecutor class is part of the 'service' module.

You can also create your own function executor. The return object of ChatFunctionCall.getArguments() is a JsonNode for simplicity and should be able to help you with that.

For a more in-depth look, refer to a conversational example that employs functions in: Or for an example using functions and stream:

Streaming thread shutdown

If you want to shut down your process immediately after streaming responses, call OpenAiService.shutdownExecutor().
This is not necessary for non-streaming calls.

Running the example project

All the example project requires is your OpenAI api token


You can try all the capabilities of this project using:

./gradlew runExampleOne

And you can also try the new capability of using functions:

./gradlew runExampleTwo

Or functions with 'stream' mode enabled:

./gradlew runExampleThree


Does this support GPT-4?

Yes! GPT-4 uses the ChatCompletion Api, and you can see the latest model options here.
GPT-4 is currently in a limited beta (as of 4/1/23), so make sure you have access before trying to use it.

Does this support functions?

Absolutely! It is very easy to use your own functions without worrying about doing the dirty work. As mentioned above, you can refer to or projects for an example.

Why am I getting connection timeouts?

Make sure that OpenAI is available in your country.

Why doesn't OpenAiService support x configuration option?

Many projects use OpenAiService, and in order to support them best I've kept it extremely simple.
You can create your own OpenAiApi instance to customize headers, timeouts, base urls etc.
If you want features like retry logic and async calls, you'll have to make an OpenAiApi instance and call it directly instead of using OpenAiService

Deprecated Endpoints

OpenAI has deprecated engine-based endpoints in favor of model-based endpoints. For example, instead of using v1/engines/{engine_id}/completions, switch to v1/completions and specify the model in the CompletionRequest. The code includes upgrade instructions for all deprecated endpoints.

I won't remove the old endpoints from this library until OpenAI shuts them down.


Published under the MIT License

About Me
GitHub: Trinea
Facebook: Dev Tools