To call the OpenAI REST API, you will need an API key. To obtain one, first create a new OpenAI account or log in. Next, navigate to the API key page and select "Create new secret key", optionally naming the key. Make sure to save your API key somewhere safe and do not share it with anyone.
Install the NuGet package
Add the client library to your .NET project by installing the NuGet package via your IDE or by running the following command in the .NET CLI:
dotnet add package OpenAI
If you would like to try the latest preview version, remember to append the --prerelease command option.
Note that the code examples included below were written using .NET 8. The OpenAI .NET library is compatible with all .NET Standard 2.0 applications, but the syntax used in some of the code examples in this document may depend on newer language features.
Using the client library
The full API of this library can be found in the OpenAI.netstandard2.0.cs file, and there are many code examples to help. For instance, the following snippet illustrates the basic use of the chat completions API:
using OpenAI.Chat;
ChatClient client = new(model: "gpt-4o", apiKey: Environment.GetEnvironmentVariable("OPENAI_API_KEY"));
ChatCompletion completion = client.CompleteChat("Say 'this is a test.'");
Console.WriteLine($"[ASSISTANT]: {completion.Content[0].Text}");
While you can pass your API key directly as a string, it is highly recommended that you keep it in a secure location and instead access it via an environment variable or configuration file as shown above to avoid storing it in source control.
Namespace organization
The library is organized into namespaces by feature areas in the OpenAI REST API. Each namespace contains a corresponding client class.
Namespace
Client class
Notes
OpenAI.Assistants
AssistantClient
OpenAI.Audio
AudioClient
OpenAI.Batch
BatchClient
OpenAI.Chat
ChatClient
OpenAI.Embeddings
EmbeddingClient
OpenAI.FineTuning
FineTuningClient
OpenAI.Files
OpenAIFileClient
OpenAI.Images
ImageClient
OpenAI.Models
OpenAIModelClient
OpenAI.Moderations
ModerationClient
OpenAI.VectorStores
VectorStoreClient
Using the async API
Every client method that performs a synchronous API call has an asynchronous variant in the same client class. For instance, the asynchronous variant of the ChatClient's CompleteChat method is CompleteChatAsync. To rewrite the call above using the asynchronous counterpart, simply await the call to the corresponding async variant:
ChatCompletion completion = await client.CompleteChatAsync("Say 'this is a test.'");
Using the OpenAIClient class
In addition to the namespaces mentioned above, there is also the parent OpenAI namespace itself:
using OpenAI;
This namespace contains the OpenAIClient class, which offers certain conveniences when you need to work with multiple feature area clients. Specifically, you can use an instance of this class to create instances of the other clients and have them share the same implementation details, which might be more efficient.
You can create an OpenAIClient by specifying the API key that all clients will use for authentication:
Next, to create an instance of an AudioClient, for example, you can call the OpenAIClient's GetAudioClient method by passing the OpenAI model that the AudioClient will use, just as if you were using the AudioClient constructor directly. If necessary, you can create additional clients of the same type to target different models.
When you request a chat completion, the default behavior is for the server to generate it in its entirety before sending it back in a single response. Consequently, long chat completions can require waiting for several seconds before hearing back from the server. To mitigate this, the OpenAI REST API supports the ability to stream partial results back as they are being generated, allowing you to start processing the beginning of the completion before it is finished.
The client library offers a convenient approach to working with streaming chat completions. If you wanted to re-write the example from the previous section using streaming, rather than calling the ChatClient's CompleteChat method, you would call its CompleteChatStreaming method instead:
CollectionResult<StreamingChatCompletionUpdate> completionUpdates = client.CompleteChatStreaming("Say 'this is a test.'");
Notice that the returned value is a CollectionResult<StreamingChatCompletionUpdate> instance, which can be enumerated to process the streaming response chunks as they arrive:
Console.Write($"[ASSISTANT]: ");
foreach (StreamingChatCompletionUpdate completionUpdate in completionUpdates)
{
if (completionUpdate.ContentUpdate.Count > 0)
{
Console.Write(completionUpdate.ContentUpdate[0].Text);
}
}
Alternatively, you can do this asynchronously by calling the CompleteChatStreamingAsync method to get an AsyncCollectionResult<StreamingChatCompletionUpdate> and enumerate it using await foreach:
AsyncCollectionResult<StreamingChatCompletionUpdate> completionUpdates = client.CompleteChatStreamingAsync("Say 'this is a test.'");
Console.Write($"[ASSISTANT]: ");
await foreach (StreamingChatCompletionUpdate completionUpdate in completionUpdates)
{
if (completionUpdate.ContentUpdate.Count > 0)
{
Console.Write(completionUpdate.ContentUpdate[0].Text);
}
}
How to use chat completions with tools and function calling
In this example, you have two functions. The first function can retrieve a user's current geographic location (e.g., by polling the location service APIs of the user's device), while the second function can query the weather in a given location (e.g., by making an API call to some third-party weather service). You want the model to be able to call these functions if it deems it necessary to have this information in order to respond to a user request as part of generating a chat completion. For illustrative purposes, consider the following:
private static string GetCurrentLocation()
{
// Call the location API here.
return "San Francisco";
}
private static string GetCurrentWeather(string location, string unit = "celsius")
{
// Call the weather API here.
return $"31 {unit}";
}
Start by creating two ChatTool instances using the static CreateFunctionTool method to describe each function:
private static readonly ChatTool getCurrentLocationTool = ChatTool.CreateFunctionTool(
functionName: nameof(GetCurrentLocation),
functionDescription: "Get the user's current location"
);
private static readonly ChatTool getCurrentWeatherTool = ChatTool.CreateFunctionTool(
functionName: nameof(GetCurrentWeather),
functionDescription: "Get the current weather in a given location",
functionParameters: BinaryData.FromBytes("""
{
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. Boston, MA"
},
"unit": {
"type": "string",
"enum": [ "celsius", "fahrenheit" ],
"description": "The temperature unit to use. Infer this from the specified location."
}
},
"required": [ "location" ]
}
"""u8.ToArray())
);
Next, create a ChatCompletionOptions instance and add both to its Tools property. You will pass the ChatCompletionOptions as an argument in your calls to the ChatClient's CompleteChat method.
List<ChatMessage> messages =
[
new UserChatMessage("What's the weather like today?"),
];
ChatCompletionOptions options = new()
{
Tools = { getCurrentLocationTool, getCurrentWeatherTool },
};
When the resulting ChatCompletion has a FinishReason property equal to ChatFinishReason.ToolCalls, it means that the model has determined that one or more tools must be called before the assistant can respond appropriately. In those cases, you must first call the function specified in the ChatCompletion's ToolCalls and then call the ChatClient's CompleteChat method again while passing the function's result as an additional ChatRequestToolMessage. Repeat this process as needed.
bool requiresAction;
do
{
requiresAction = false;
ChatCompletion completion = client.CompleteChat(messages, options);
switch (completion.FinishReason)
{
case ChatFinishReason.Stop:
{
// Add the assistant message to the conversation history.
messages.Add(new AssistantChatMessage(completion));
break;
}
case ChatFinishReason.ToolCalls:
{
// First, add the assistant message with tool calls to the conversation history.
messages.Add(new AssistantChatMessage(completion));
// Then, add a new tool message for each tool call that is resolved.
foreach (ChatToolCall toolCall in completion.ToolCalls)
{
switch (toolCall.FunctionName)
{
case nameof(GetCurrentLocation):
{
string toolResult = GetCurrentLocation();
messages.Add(new ToolChatMessage(toolCall.Id, toolResult));
break;
}
case nameof(GetCurrentWeather):
{
// The arguments that the model wants to use to call the function are specified as a
// stringified JSON object based on the schema defined in the tool definition. Note that
// the model may hallucinate arguments too. Consequently, it is important to do the
// appropriate parsing and validation before calling the function.
using JsonDocument argumentsJson = JsonDocument.Parse(toolCall.FunctionArguments);
bool hasLocation = argumentsJson.RootElement.TryGetProperty("location", out JsonElement location);
bool hasUnit = argumentsJson.RootElement.TryGetProperty("unit", out JsonElement unit);
if (!hasLocation)
{
throw new ArgumentNullException(nameof(location), "The location argument is required.");
}
string toolResult = hasUnit
? GetCurrentWeather(location.GetString(), unit.GetString())
: GetCurrentWeather(location.GetString());
messages.Add(new ToolChatMessage(toolCall.Id, toolResult));
break;
}
default:
{
// Handle other unexpected calls.
throw new NotImplementedException();
}
}
}
requiresAction = true;
break;
}
case ChatFinishReason.Length:
throw new NotImplementedException("Incomplete model output due to MaxTokens parameter or token limit exceeded.");
case ChatFinishReason.ContentFilter:
throw new NotImplementedException("Omitted content due to a content filter flag.");
case ChatFinishReason.FunctionCall:
throw new NotImplementedException("Deprecated in favor of tool calls.");
default:
throw new NotImplementedException(completion.FinishReason.ToString());
}
} while (requiresAction);
How to use chat completions with structured outputs
Beginning with the gpt-4o-mini, gpt-4o-mini-2024-07-18, and gpt-4o-2024-08-06 model snapshots, structured outputs are available for both top-level response content and tool calls in the chat completion and assistants APIs. For information about the feature, see the Structured Outputs guide.
To use structured outputs to constrain chat completion content, set an appropriate ChatResponseFormat as in the following example:
In this example, you want to create a trip-planning website that allows customers to write a prompt describing the kind of hotel that they are looking for and then offers hotel recommendations that closely match this description. To achieve this, it is possible to use text embeddings to measure the relatedness of text strings. In summary, you can get embeddings of the hotel descriptions, store them in a vector database, and use them to build a search index that you can query using the embedding of a given customer's prompt.
To generate a text embedding, use EmbeddingClient from the OpenAI.Embeddings namespace:
using OpenAI.Embeddings;
EmbeddingClient client = new("text-embedding-3-small", Environment.GetEnvironmentVariable("OPENAI_API_KEY"));
string description = "Best hotel in town if you like luxury hotels. They have an amazing infinity pool, a spa,"
+ " and a really helpful concierge. The location is perfect -- right downtown, close to all the tourist"
+ " attractions. We highly recommend this hotel.";
OpenAIEmbedding embedding = client.GenerateEmbedding(description);
ReadOnlyMemory<float> vector = embedding.ToFloats();
Notice that the resulting embedding is a list (also called a vector) of floating point numbers represented as an instance of ReadOnlyMemory<float>. By default, the length of the embedding vector will be 1536 when using the text-embedding-3-small model or 3072 when using the text-embedding-3-large model. Generally, larger embeddings perform better, but using them also tends to cost more in terms of compute, memory, and storage. You can reduce the dimensions of the embedding by creating an instance of the EmbeddingGenerationOptions class, setting the Dimensions property, and passing it as an argument in your call to the GenerateEmbedding method:
In this example, you want to build an app to help interior designers prototype new ideas based on the latest design trends. As part of the creative process, an interior designer can use this app to generate images for inspiration simply by describing the scene in their head as a prompt. As expected, high-quality, strikingly dramatic images with finer details deliver the best results for this application.
To generate an image, use ImageClient from the OpenAI.Images namespace:
using OpenAI.Images;
ImageClient client = new("dall-e-3", Environment.GetEnvironmentVariable("OPENAI_API_KEY"));
Generating an image always requires a prompt that describes what should be generated. To further tailor the image generation to your specific needs, you can create an instance of the ImageGenerationOptions class and set the Quality, Size, and Style properties accordingly. Note that you can also set the ResponseFormat property of ImageGenerationOptions to GeneratedImageFormat.Bytes in order to receive the resulting PNG as BinaryData (instead of the default remote Uri) if this is convenient for your use case.
string prompt = "The concept for a living room that blends Scandinavian simplicity with Japanese minimalism for"
+ " a serene and cozy atmosphere. It's a space that invites relaxation and mindfulness, with natural light"
+ " and fresh air. Using neutral tones, including colors like white, beige, gray, and black, that create a"
+ " sense of harmony. Featuring sleek wood furniture with clean lines and subtle curves to add warmth and"
+ " elegance. Plants and flowers in ceramic pots adding color and life to a space. They can serve as focal"
+ " points, creating a connection with nature. Soft textiles and cushions in organic fabrics adding comfort"
+ " and softness to a space. They can serve as accents, adding contrast and texture.";
ImageGenerationOptions options = new()
{
Quality = GeneratedImageQuality.High,
Size = GeneratedImageSize.W1792xH1024,
Style = GeneratedImageStyle.Vivid,
ResponseFormat = GeneratedImageFormat.Bytes
};
Finally, call the ImageClient's GenerateImage method by passing the prompt and the ImageGenerationOptions instance as arguments:
For illustrative purposes, you could then save the generated image to local storage:
using FileStream stream = File.OpenWrite($"{Guid.NewGuid()}.png");
bytes.ToStream().CopyTo(stream);
How to transcribe audio
In this example, an audio file is transcribed using the Whisper speech-to-text model, including both word- and audio-segment-level timestamp information.
How to use assistants with retrieval augmented generation (RAG)
In this example, you have a JSON document with the monthly sales information of different products, and you want to build an assistant capable of analyzing it and answering questions about it.
To achieve this, use both OpenAIFileClient from the OpenAI.Files namespace and AssistantClient from the OpenAI.Assistants namespace.
Important: The Assistants REST API is currently in beta. As such, the details are subject to change, and correspondingly the AssistantClient is attributed as [Experimental]. To use it, you must suppress the OPENAI001 warning first.
using OpenAI.Assistants;
using OpenAI.Files;
OpenAIClient openAIClient = new(Environment.GetEnvironmentVariable("OPENAI_API_KEY"));
OpenAIFileClient fileClient = openAIClient.GetOpenAIFileClient();
AssistantClient assistantClient = openAIClient.GetAssistantClient();
Here is an example of what the JSON document might look like:
using Stream document = BinaryData.FromBytes("""
{
"description": "This document contains the sale history data for Contoso products.",
"sales": [
{
"month": "January",
"by_product": {
"113043": 15,
"113045": 12,
"113049": 2
}
},
{
"month": "February",
"by_product": {
"113045": 22
}
},
{
"month": "March",
"by_product": {
"113045": 16,
"113055": 5
}
}
]
}
"""u8.ToArray()).ToStream();
Upload this document to OpenAI using the OpenAIFileClient's UploadFile method, ensuring that you use FileUploadPurpose.Assistants to allow your assistant to access it later:
Create a new assistant using an instance of the AssistantCreationOptions class to customize it. Here, we use:
A friendly Name for the assistant, as will display in the Playground
Tool definition instances for the tools that the assistant should have access to; here, we use FileSearchToolDefinition to process the sales document we just uploaded and CodeInterpreterToolDefinition so we can analyze and visualize the numeric data
Resources for the assistant to use with its tools, here using the VectorStoreCreationHelper type to automatically make a new vector store that indexes the sales file; alternatively, you could use VectorStoreClient to manage the vector store separately
AssistantCreationOptions assistantOptions = new()
{
Name = "Example: Contoso sales RAG",
Instructions =
"You are an assistant that looks up sales data and helps visualize the information based"
+ " on user queries. When asked to generate a graph, chart, or other visualization, use"
+ " the code interpreter tool to do so.",
Tools =
{
new FileSearchToolDefinition(),
new CodeInterpreterToolDefinition(),
},
ToolResources = new()
{
FileSearch = new()
{
NewVectorStores =
{
new VectorStoreCreationHelper([salesFile.Id]),
}
}
},
};
Assistant assistant = assistantClient.CreateAssistant("gpt-4o", assistantOptions);
Next, create a new thread. For illustrative purposes, you could include an initial user message asking about the sales information of a given product and then use the AssistantClient's CreateThreadAndRun method to get it started:
ThreadCreationOptions threadOptions = new()
{
InitialMessages = { "How well did product 113045 sell in February? Graph its trend over time." }
};
ThreadRun threadRun = assistantClient.CreateThreadAndRun(assistant.Id, threadOptions);
Poll the status of the run until it is no longer queued or in progress:
do
{
Thread.Sleep(TimeSpan.FromSeconds(1));
threadRun = assistantClient.GetRun(threadRun.ThreadId, threadRun.Id);
} while (!threadRun.Status.IsTerminal);
If everything went well, the terminal status of the run will be RunStatus.Completed.
Finally, you can use the AssistantClient's GetMessages method to retrieve the messages associated with this thread, which now include the responses from the assistant to the initial user message.
For illustrative purposes, you could print the messages to the console and also save any images produced by the assistant to local storage:
CollectionResult<ThreadMessage> messages
= assistantClient.GetMessages(threadRun.ThreadId, new MessageCollectionOptions() { Order = MessageCollectionOrder.Ascending });
foreach (ThreadMessage message in messages)
{
Console.Write($"[{message.Role.ToString().ToUpper()}]: ");
foreach (MessageContent contentItem in message.Content)
{
if (!string.IsNullOrEmpty(contentItem.Text))
{
Console.WriteLine($"{contentItem.Text}");
if (contentItem.TextAnnotations.Count > 0)
{
Console.WriteLine();
}
// Include annotations, if any.
foreach (TextAnnotation annotation in contentItem.TextAnnotations)
{
if (!string.IsNullOrEmpty(annotation.InputFileId))
{
Console.WriteLine($"* File citation, file ID: {annotation.InputFileId}");
}
if (!string.IsNullOrEmpty(annotation.OutputFileId))
{
Console.WriteLine($"* File output, new file ID: {annotation.OutputFileId}");
}
}
}
if (!string.IsNullOrEmpty(contentItem.ImageFileId))
{
OpenAIFile imageInfo = fileClient.GetFile(contentItem.ImageFileId);
BinaryData imageBytes = fileClient.DownloadFile(contentItem.ImageFileId);
using FileStream stream = File.OpenWrite($"{imageInfo.Filename}.png");
imageBytes.ToStream().CopyTo(stream);
Console.WriteLine($"<image: {imageInfo.Filename}.png>");
}
}
Console.WriteLine();
}
And it would yield something like this:
[USER]: How well did product 113045 sell in February? Graph its trend over time.
[ASSISTANT]: Product 113045 sold 22 units in February【4:0†monthly_sales.json】.
Now, I will generate a graph to show its sales trend over time.
* File citation, file ID: file-hGOiwGNftMgOsjbynBpMCPFn
[ASSISTANT]: <image: 015d8e43-17fe-47de-af40-280f25452280.png>
The sales trend for Product 113045 over the past three months shows that:
- In January, 12 units were sold.
- In February, 22 units were sold, indicating significant growth.
- In March, sales dropped slightly to 16 units.
The graph above visualizes this trend, showing a peak in sales during February.
How to use assistants with streaming and vision
This example shows how to use the v2 Assistants API to provide image data to an assistant and then stream the run's response.
As before, you will use a OpenAIFileClient and an AssistantClient:
For this example, we will use both image data from a local file as well as an image located at a URL. For the local data, we upload the file with the Vision upload purpose, which would also allow it to be downloaded and retrieved later.
Next, create a new assistant with a vision-capable model like gpt-4o and a thread with the image information referenced:
Assistant assistant = assistantClient.CreateAssistant(
"gpt-4o",
new AssistantCreationOptions()
{
Instructions = "When asked a question, attempt to answer very concisely. "
+ "Prefer one-sentence answers whenever feasible."
});
AssistantThread thread = assistantClient.CreateThread(new ThreadCreationOptions()
{
InitialMessages =
{
new ThreadInitializationMessage(
MessageRole.User,
[
"Hello, assistant! Please compare these two images for me:",
MessageContent.FromImageFileId(pictureOfAppleFile.Id),
MessageContent.FromImageUri(linkToPictureOfOrange),
]),
}
});
With the assistant and thread prepared, use the CreateRunStreaming method to get an enumerable CollectionResult<StreamingUpdate>. You can then iterate over this collection with foreach. For async calling patterns, use CreateRunStreamingAsync and iterate over the AsyncCollectionResult<StreamingUpdate> with await foreach, instead. Note that streaming variants also exist for CreateThreadAndRunStreaming and SubmitToolOutputsToRunStreaming.
CollectionResult<StreamingUpdate> streamingUpdates = assistantClient.CreateRunStreaming(
thread.Id,
assistant.Id,
new RunCreationOptions()
{
AdditionalInstructions = "When possible, try to sneak in puns if you're asked to compare things.",
});
Finally, to handle the StreamingUpdates as they arrive, you can use the UpdateKind property on the base StreamingUpdate and/or downcast to a specifically desired update type, like MessageContentUpdate for thread.message.delta events or RequiredActionUpdate for streaming tool calls.
foreach (StreamingUpdate streamingUpdate in streamingUpdates)
{
if (streamingUpdate.UpdateKind == StreamingUpdateReason.RunCreated)
{
Console.WriteLine($"--- Run started! ---");
}
if (streamingUpdate is MessageContentUpdate contentUpdate)
{
Console.Write(contentUpdate.Text);
}
}
This will yield streamed output from the run like the following:
--- Run started! ---
The first image depicts a multicolored apple with a blend of red and green hues, while the second image shows an orange with a bright, textured orange peel; one might say it’s comparing apples to oranges!
The Azure OpenAI client library for .NET is a companion to this library and all common capabilities between OpenAI and Azure OpenAI share the same scenario clients, methods, and request/response types. It is designed to make Azure specific scenarios straightforward, with extensions for Azure-specific concepts like Responsible AI content filter results and On Your Data integration.
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
ChatClient chatClient = azureClient.GetChatClient("my-gpt-35-turbo-deployment");
ChatCompletion completion = chatClient.CompleteChat(
[
// System messages represent instructions or other guidance about how the assistant should behave
new SystemChatMessage("You are a helpful assistant that talks like a pirate."),
// User messages represent user input, whether historical or the most recen tinput
new UserChatMessage("Hi, can you help me?"),
// Assistant messages in a request represent conversation history for responses
new AssistantChatMessage("Arrr! Of course, me hearty! What can I do for ye?"),
new UserChatMessage("What's the best way to train a parrot?"),
]);
Console.WriteLine($"{completion.Role}: {completion.Content[0].Text}");
Advanced scenarios
Using protocol methods
In addition to the client methods that use strongly-typed request and response objects, the .NET library also provides protocol methods that enable more direct access to the REST API. Protocol methods are "binary in, binary out" accepting BinaryContent as request bodies and providing BinaryData as response bodies.
For example, to use the protocol method variant of the ChatClient's CompleteChat method, pass the request body as BinaryContent:
Notice how you can then call the resulting ClientResult's GetRawResponse method and retrieve the response body as BinaryData via the PipelineResponse's Content property.
Mock a client for testing
The OpenAI .NET library has been designed to support mocking, providing key features such as:
Client methods made virtual to allow overriding.
Model factories to assist in instantiating API output models that lack public constructors.
To illustrate how mocking works, suppose you want to validate the behavior of the following method using the Moq library. Given the path to an audio file, it determines whether it contains a specified secret word:
Create mocks of AudioClient and ClientResult<AudioTranscription>, set up methods and properties that will be invoked, then test the behavior of the ContainsSecretWord method. Since the AudioTranscription class does not provide public constructors, it must be instantiated by the OpenAIAudioModelFactory static class:
// Instantiate mocks and the AudioTranscription object.
Mock<AudioClient> mockClient = new();
Mock<ClientResult<AudioTranscription>> mockResult = new(null, Mock.Of<PipelineResponse>());
AudioTranscription transcription = OpenAIAudioModelFactory.AudioTranscription(text: "I swear I saw an apple flying yesterday!");
// Set up mocks' properties and methods.
mockResult
.SetupGet(result => result.Value)
.Returns(transcription);
mockClient.Setup(client => client.TranscribeAudio(
It.IsAny<string>(),
It.IsAny<AudioTranscriptionOptions>()))
.Returns(mockResult.Object);
// Perform validation.
AudioClient client = mockClient.Object;
bool containsSecretWord = ContainsSecretWord(client, "<audioFilePath>", "apple");
Assert.That(containsSecretWord, Is.True);
All namespaces have their corresponding model factory to support mocking with the exception of the OpenAI.Assistants and OpenAI.VectorStores namespaces, for which model factories are coming soon.
Automatically retrying errors
By default, the client classes will automatically retry the following errors up to three additional times using exponential backoff:
408 Request Timeout
429 Too Many Requests
500 Internal Server Error
502 Bad Gateway
503 Service Unavailable
504 Gateway Timeout
Observability
OpenAI .NET library supports experimental distributed tracing and metrics with OpenTelemetry. Check out Observability with OpenTelemetry for more details.
