A ChatGPT integration library for .NET, supporting both OpenAI and Azure OpenAI Service.
The library is available on NuGet. Just search for ChatGptNet in the Package Manager GUI or run the following command in the .NET CLI:
dotnet add package ChatGptNet
Register ChatGPT service at application startup:
builder.Services.AddChatGpt(options =>
{
// OpenAI.
//options.UseOpenAI(apiKey: "", organization: "");
// Azure OpenAI Service.
//options.UseAzure(resourceName: "", apiKey: "", authenticationType: AzureAuthenticationType.ApiKey);
options.DefaultModel = "my-model";
options.MessageLimit = 16; // Default: 10
options.MessageExpiration = TimeSpan.FromMinutes(5); // Default: 1 hour
});
ChatGptNet supports both OpenAI and Azure OpenAI Service, so it is necessary to set the correct configuration settings based on the chosen provider:
- ApiKey: it is available in the User settings page of the OpenAI account (required).
- Organization: for users who belong to multiple organizations, you can also specify which organization is used. Usage from these API requests will count against the specified organization's subscription quota (optional).
- ResourceName: the name of your Azure OpenAI Resource (required).
- ApiKey: Azure OpenAI provides two methods for authentication. You can use either API Keys or Azure Active Directory (required).
- ApiVersion: the version of the API to use (optional). Allowed values:
- 2023-03-15-preview
- 2023-05-15
- 2023-06-01-preview
- 2023-07-01-preview (default)
- AuthenticationType: it specifies if the key is an actual API Key or an Azure Active Directory token (optional, default: "ApiKey").
ChatGPT can be used with different models for chat completion, both on OpenAI and Azure OpenAI service. With the DefaultModel property, you can specify the default model that will be used, unless you pass an explicit value in the AskAsync method.
Currently available models are: gpt-3.5-turbo, gpt-3.5-turbo-16k, gpt-4 and gpt-4-32k. They have fixed names, available in the OpenAIChatGptModels.cs file.
In Azure OpenAI Service, you're required to first deploy a model before you can make calls. When you deploy a model, you need to assign it a name, that must match the name you use with ChatGptNet.
Note Some models are not available in all regions. You can refer to Model Summary table and region availability page to check current availabilities.
ChatGPT is aimed to support conversational scenarios: user can talk to ChatGPT without specifying the full context for every interaction. However, conversation history isn't managed by OpenAI or Azure OpenAI service, so it's up to us to retain the current state. By default, ChatGptNet handles this requirement using a MemoryCache that stores messages for each conversation. The behavior can be set using the following properties:
- MessageLimit: specifies how many messages for each conversation must be saved. When this limit is reached, oldest messages are automatically removed.
- MessageExpiration: specifies the time interval used to maintain messages in cache, regardless their count.
If necessary, it is possibile to provide a custom Cache by implementing the IChatGptCache interface and then calling the WithCache extension method:
public class LocalMessageCache : IChatGptCache
{
private readonly Dictionary<Guid, List<ChatGptMessage>> localCache = new();
public Task SetAsync(Guid conversationId, IEnumerable<ChatGptMessage> messages, TimeSpan expiration, CancellationToken cancellationToken = default)
{
localCache[conversationId] = messages.ToList();
return Task.CompletedTask;
}
public Task<List<ChatGptMessage>?> GetAsync(Guid conversationId, CancellationToken cancellationToken = default)
{
localCache.TryGetValue(conversationId, out var messages);
return Task.FromResult(messages);
}
public Task RemoveAsync(Guid conversationId, CancellationToken cancellationToken = default)
{
localCache.Remove(conversationId);
return Task.CompletedTask;
}
public Task<bool> ExistsAsync(Guid conversationId, CancellationToken cancellationToken = default)
{
var exists = localCache.ContainsKey(conversationId);
return Task.FromResult(exists);
}
}
// Registers the custom cache at application startup.
builder.Services.AddChatGpt(/* ... */).WithCache<LocalMessageCache>();
We can also set ChatGPT parameters for chat completion at startup. Check the official documentation for the list of available parameters and their meaning.
The configuration can be automatically read from IConfiguration, using for example a ChatGPT section in the appsettings.json file:
"ChatGPT": {
"Provider": "OpenAI", // Optional. Allowed values: OpenAI (default) or Azure
"ApiKey": "", // Required
//"Organization": "", // Optional, used only by OpenAI
"ResourceName": "", // Required when using Azure OpenAI Service
"ApiVersion": "2023-07-01-preview", // Optional, used only by Azure OpenAI Service (default: 2023-07-01-preview)
"AuthenticationType": "ApiKey", // Optional, used only by Azure OpenAI Service. Allowed values: ApiKey (default) or ActiveDirectory
"DefaultModel": "my-model",
"MessageLimit": 20,
"MessageExpiration": "00:30:00",
"ThrowExceptionOnError": true
//"User": "UserName",
//"DefaultParameters": {
// "Temperature": 0.8,
// "TopP": 1,
// "MaxTokens": 500,
// "PresencePenalty": 0,
// "FrequencyPenalty": 0
//}
}
And then use the corresponding overload of che AddChatGpt method:
// Adds ChatGPT service using settings from IConfiguration.
builder.Services.AddChatGpt(builder.Configuration);
The AddChatGpt method has also an overload that accepts an IServiceProvider as argument. It can be used, for example, if we're in a Web API and we need to support scenarios in which every user has a different API Key that can be retrieved accessing a database via Dependency Injection:
builder.Services.AddChatGpt((services, options) =>
{
var accountService = services.GetRequiredService<IAccountService>();
// Dynamically gets the API Key from the service.
var apiKey = "..."
options.UseOpenAI(apiKyey);
});
In more complex scenarios, it is possible to configure ChatGptNet using both code and IConfiguration. This can be useful if we want to set a bunch of common properties, but at the same time we need some configuration logic. For example:
builder.Services.AddChatGpt((services, options) =>
{
// Configure common properties (message limit and expiration, default parameters, ecc.) using IConfiguration.
options.UseConfiguration(builder.Configuration);
var accountService = services.GetRequiredService<IAccountService>();
// Dynamically gets the API Key from the service.
var apiKey = "..."
options.UseOpenAI(apiKyey);
});
The library can be used in any .NET application built with .NET 6.0 or later. For example, we can create a Minimal API in this way:
app.MapPost("/api/chat/ask", async (Request request, IChatGptClient chatGptClient) =>
{
var response = await chatGptClient.AskAsync(request.ConversationId, request.Message);
return TypedResults.Ok(response);
})
.WithOpenApi();
// ...
public record class Request(Guid ConversationId, string Message);
If we just want to retrieve the response message, we can call the GetMessage method:
var message = response.GetMessage();
The AskAsync method has an overload (the one shown in the example above) that requires a conversationId parameter. If we pass an empty value, a random one is generated and returned. We can pass this value in subsequent invocations of AskAsync so that the library automatically retrieves previous messages of the current conversation (according to MessageLimit and MessageExpiration settings) and send them to chat completion API.
Chat completion API supports response streaming. When using this feature, partial message deltas will be sent, like in ChatGPT. Tokens will be sent as data-only server-sent events as they become available. ChatGptNet provides response streaming using the AskStreamAsync method:
// Requests a streaming response.
var responseStream = chatGptClient.AskStreamAsync(conversationId, message);
await foreach (var response in responseStream)
{
Console.Write(response.GetMessage());
await Task.Delay(80);
}
Response streaming works by returning an IAsyncEnumerable, so it can be used even in a Web API project:
app.MapGet("/api/chat/stream", (Guid? conversationId, string message, IChatGptClient chatGptClient) =>
{
async IAsyncEnumerable<string> Stream()
{
// Requests a streaming response.
var responseStream = chatGptClient.AskStreamAsync(conversationId.GetValueOrDefault(), message);
// Uses the "AsDeltas" extension method to retrieve partial message deltas only.
await foreach (var delta in responseStream.AsDeltas())
{
yield return delta;
await Task.Delay(50);
}
}
return Stream();
})
.WithOpenApi();
The library is 100% compatible also with Blazor WebAssembly applications:
Check out the Samples folder for more information about the different implementations.
ChatGPT supports messages with the system role to influence how the assistant should behave. For example, we can tell to ChatGPT something like that:
- You are an helpful assistant
- Answer like Shakespeare
- Give me only wrong answers
- Answer in rhyme
ChatGptNet provides this feature using the SetupAsync method:
var conversationId await = chatGptClient.SetupAsync("Answer in rhyme");
If we use the same conversationId when calling AskAsync, then the system message will be automatically sent along with every request, so that the assistant will know how to behave.
Note The system message does not count for messages limit number.
Conversation history is automatically deleted when expiration time (specified by MessageExpiration property) is reached. However, if necessary it is possible to immediately clear the history:
await chatGptClient.DeleteConversationAsync(conversationId, preserveSetup: false);
The preserveSetup argument allows to decide whether mantain also the system message that has been set with the SetupAsync method (default: false).
With function calling, we can describe functions and have the model intelligently choose to output a JSON object containing arguments to call those functions. This is a new way to more reliably connect GPT's capabilities with external tools and APIs.
Note Currently, on Azure OpenAI Service, function calling is supported in the following models in API version
2023-07-01-preview:
- gpt-35-turbo-0613
- gpt-35-turbo-16k-0613
- gpt-4-0613
- gpt-4-32k-0613
ChatGptNet fully supports function calling by providing an overload of the AskAsync method that allows to specify function definitions. If this parameter is supplied, then the model will decide when it is appropiate to use one the functions. For example:
var functions = new List<ChatGptFunction>
{
new()
{
Name = "GetCurrentWeather",
Description = "Get the current weather",
Parameters = JsonDocument.Parse("""
{
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and/or the zip code"
},
"format": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location."
}
},
"required": ["location", "format"]
}
""")
},
new()
{
Name = "GetWeatherForecast",
Description = "Get an N-day weather forecast",
Parameters = JsonDocument.Parse("""
{
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and/or the zip code"
},
"format": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "The temperature unit to use. Infer this from the users location."
},
"daysNumber": {
"type": "integer",
"description": "The number of days to forecast"
}
},
"required": ["location", "format", "daysNumber"]
}
""")
}
};
var functionParameters = new ChatGptFunctionParameters
{
FunctionCall = ChatGptFunctionCalls.Auto, // This is the default if functions are present.
Functions = functions
};
var response = await chatGptClient.AskAsync("What is the weather like in Taggia?", functionParameters);
We can pass an arbitrary number of functions, each one with a name, a description and a JSON schema describing the function parameters, following the JSON Schema references. Under the hood, functions are injected into the system message in a syntax the model has been trained on. This means functions count against the model's context limit and are billed as input tokens.
The response object returned by the AskAsync method provides a property to check if the model has selected a function call:
if (response.IsFunctionCall)
{
Console.WriteLine("I have identified a function to call:");
var functionCall = response.GetFunctionCall()!;
Console.WriteLine(functionCall.Name);
Console.WriteLine(functionCall.Arguments);
}
This code will print something like this:
I have identified a function to call:
GetCurrentWeather
{
"location": "Taggia",
"format": "celsius"
}
Note that the API will not actually execute any function calls. It is up to developers to execute function calls using model outputs.
After the actual execution, we need to call the AddFunctionResponseAsync method on the ChatGptClient to add the response to the conversation history, just like a standard message, so that it will be automatically used for chat completion:
// Calls the remote function API.
var functionResponse = await GetWeatherAsync(functionCall.Arguments);
await chatGptClient.AddFunctionResponseAsync(conversationId, functionCall.Name, functionResponse);
Check out the Function calling sample for a complete implementation of this workflow.
The project is constantly evolving. Contributions are welcome. Feel free to file issues and pull requests on the repo and we'll address them as we can.
Warning Remember to work on the develop branch, don't use the master branch directly. Create Pull Requests targeting develop.