OpenAI Chat Completions API (ChatGPT) integration with DI and EF Core support. It allows you to use the API in your .NET applications. Also, the client supports streaming responses (like ChatGPT) via async streams.
- Add OpenRouter and Azure OpenAI support
- Add support for GPT-4-o, GPT-4-o-mini
- Improve JSON mode support
- BREAKING CHANGES:
AddChatGptEntityFrameworkIntegration
now requiresbuilder.Configuration
as a mandatory parameter. - Add a setup example
- Add GPT-4-Turbo
- Add JSON mode support
StructuredResponse
module allows you to get structured responses from the API as C# object. See: StructuredResponse section.
First, you need to create an OpenAI account and get an API key (or OpenRouter or Azure OpenAI). You can do this at https://platform.openai.com/account/api-keys.
The easiest way to use ChatGPT service in your .NET project with DI and persistence (EF Core) supporting is to install the NuGet package OpenAI.ChatGPT.EntityFrameworkCore:
Install-Package OpenAI.ChatGPT.EntityFrameworkCore
If you don't want to use EF Core, you can install the package OpenAI.ChatGPT.AspNetCore and implement your own storage for chat history, using IChatHistoryStorage
interface.
TL;DR: See an example in a reference project.
- Set the OpenAI API key or even host (optional) in your project user secrets, or the
appsettings.json
file (not safe):
{
"AIProvider": "openai", // or openrouter or azure_openai
"OpenAICredentials": { //optional
"ApiKey": "your-api-key-from-openai",
"ApiHost": "https://api.openai.com/v1/"
},
"AzureOpenAICredentials": { //optional
"ApiKey": "your-api-key-from-azure-openai",
"ApiHost": "https://{your-host}.openai.azure.com/",
"DeploymentName": "gpt-4-turbo-preview"
},
"OpenRouterCredentials": { //optional
"ApiKey": "your-api-key-from-openrouter",
"ApiHost": "https://openrouter.ai/api/v1"
}
}
Also, you can specify OpenAI API key as environment variable ASPNETCORE_OpenAICredentials:ApiKey
.
- Add ChatGPT integration with EF to your DI container:
builder.Services.AddChatGptEntityFrameworkIntegration(
builder.Configuration,
options => options.UseSqlite("Data Source=chats.db"));
Instead of options.UseSqlite("Data Source=chats.db")
use your own db and connection string.
- Inject
ChatGPTFactory
to your service and use it to createChatGPT
instance:
public class YourService
{
private readonly ChatGPTFactory _chatGptFactory;
public YourService(ChatGPTFactory chatGptFactory)
{
_chatGptFactory = chatGptFactory;
}
public async Task<string> GetAnswer(string text)
{
ChatGPT chatGpt = await _chatGptFactory.Create(userId);
var chatService = await chatGpt.ContinueOrStartNewTopic();
response = await _chatService.GetNextMessageResponse(_prompt);
return response;
}
}
See Blazor Example.
If you want to configure request parameters, you can do it in appsettings.json
configuration or in ChatGPTFactory.Create
or in ChatGPT.CreateChat
methods.
{
"ChatGPTConfig": {
"InitialSystemMessage": "You are a helpful and kind assistant.",
"InitialUserMessage": null,
"MaxTokens": null,
"Model": null,
"Temperature": null,
"PassUserIdToOpenAiRequests": true
}
}
See parameters description inside ChatGPTConfig.
If the server response is not a success status code, the client will throw a NotExpectedResponseException. The exception will contain the error message from the OpenAI API.
By default, requesting cancellation or ChatService.Stop()
method calling will throw OperationCanceledException
. If you don't want to throw it (relevant for streaming responses), you can set throwOnCancellation
parameter to false
:
await foreach (string chunk in chatService.StreamNextMessageResponse(text, throwOnCancellation: false))
{
//...
}
ChatGPTFactory
, ChatGPT
classes thread-safety is depend on the IChatHistoryStorage
implementation. If you use ChatGPTFactory
with entity framework, it's NOT thread-safe. ChatService
class is not thread-safe.
Anyways, these services are designed to be used safely with DI, so you don't need to worry about it.
All the methods from all the packages are designed to be used in async context and use ConfigureAwait(false)
(thanks for the ConfigureAwait.Fody
package).
Since ChatGPTFactory
depends on IHttpClientFactory
, you can easily use any of the available policies for it, like Polly
.
This module allows you to get structured responses from the API as C# object. It's useful if you want to use the API for something more than just chat.
record City(string Name, int YearOfFoundation, string Country);
var message = Dialog
.StartAsSystem("Return requested data.")
.ThenUser("I need info about Almaty city");
City almaty = await _client.GetStructuredResponse<City>(message, model: ChatCompletionModels.Gpt4Turbo);
Console.WriteLine(almaty); // Name: "Almaty", Country: "Kazakhstan", YearOfFoundation: 1854
Under the hood, it uses the new json mode of the API for GPT4Turbo and for the gpt-3.5-turbo-1106
. Regular GPT4 and GPT3.5Turbo models are also supported, but GPT3.5 responses may be unstable (for GPT3.5 it's strictly recommended to provide examples
parameter).
More complex examples with arrays, nested objects and enums are available in tests:
NuGet: https://www.nuget.org/packages/OpenAI.ChatGPT.Modules.StructuredResponse
This module allows you to translate messages from one language to another.
string textToTranslate = "Hello, world!";
string translatedText = await _client.TranslateText(textToTranslate, "English", "Russian");
Console.WriteLine(translatedText); // "Привет, мир!"
Also, it's possible to translate entire object in pair with StructuredResponse
module:
var objectToTranslate = new Order(
new List<Order.Item>
{
new(1,"Book", 5),
new(2,"Pen", 10),
new(3,"Notebook", 3)
}
);
Order translatedObject = await _client.TranslateObject(objectToTranslate, "English", "Russian");
See full example in tests:
NuGet: https://www.nuget.org/packages/OpenAI.ChatGPT.Modules.Translator
- Blazor Example
- Console Example (simple)
- Spectre Console Example (advanced)
Here is a list of the main parameters that can be used in the ChatCompletions (ChatGPT) API request (OpenAI.ChatGpt/Models/ChatCompletion/ChatCompletionRequest.cs).
Some of them are taken from this article: https://towardsdatascience.com/gpt-3-parameters-and-prompt-design-1a595dc5b405
Below listed parameters for ChatCompletions API.
The prediction-generating AI model is specified by the engine parameter. The available models are described below: https://platform.openai.com/docs/models
C# Model | API Model |
---|---|
ChatCompletionModels.Gpt4Turbo | gpt-4-1106-preview |
ChatCompletionModels.Gpt4 | gpt-4 |
ChatCompletionModels.Gpt4_0613 | gpt-4-0613 |
ChatCompletionModels.Gpt4_32k | gpt-4-32k |
ChatCompletionModels.Gpt4_32k_0613 | gpt-4-32k-0613 |
ChatCompletionModels.Gpt3_5_Turbo | gpt-3.5-turbo |
ChatCompletionModels.Gpt3_5_Turbo_1106 | gpt-3.5-turbo-1106 |
ChatCompletionModels.Gpt3_5_Turbo_16k | gpt-3.5-turbo-16k |
ChatCompletionModels.Gpt3_5_Turbo_0613 | gpt-3.5-turbo-0613 |
ChatCompletionModels.Gpt3_5_Turbo_16k_0613 | gpt-3.5-turbo-16k-0613 |
ChatCompletionModels.Gpt4_0314 | gpt-4-0314 |
ChatCompletionModels.Gpt4_32k_0314 | gpt-4-32k-0314 |
ChatCompletionModels.Gpt3_5_Turbo_0301 | gpt-3.5-turbo-0301 |
The maximum number of tokens allowed for the generated answer. Defaults to ChatCompletionRequest.MaxTokensDefault
(64).
- This value is validated and limited with
ChatCompletionModels.GetMaxTokensLimitForModel
method. - It's possible to calculate approximately tokens count using
ChatCompletionMessage.CalculateApproxTotalTokenCount
method - The number of tokens can be retrieved from the API response:
ChatCompletionResponse.Usage.TotalTokens
. As a rule of thumb for English, 1 token is around 4 characters (so 100 tokens ≈ 75 words). See tokenizer from OpenAI: https://platform.openai.com/tokenizer - Encoding algorithm can be found here: https://github.com/latitudegames/GPT-3-Encoder
What sampling temperature to use, between 0 and 2.
- Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.
- Predefined values are available in
ChatCompletionTemperatures
. - Default value is:
ChatCompletionTemperatures.Balanced
(0.5).
Description: Before being mapped into probabilities, the model outputs unnormalized values (logits). The logits are typically used with a function such as softmax to convert them into probabilities.
But, before applying the softmax function, we can use a trick inspired by thermodynamics and scale the logits with the temperature parameter, i.e. softmax(logits/temperature).
A temperature parameter close to 1 would mean that the logits are passed through the softmax function without modification. If the temperature is close to zero, the highest probable tokens will become very likely compared to the other tokens, i.e. the model becomes more deterministic and will always output the same set of tokens after a given sequence of words.
More parameters description can be found here: Some of them are taken from this article: https://towardsdatascience.com/gpt-3-parameters-and-prompt-design-1a595dc5b405
If you don't need DI and chat history, you can use only the NuGet package OpenAI.ChatGPT:
Install-Package OpenAI.ChatGPT
Then create an instance of OpenAIClient
:
_client = new OpenAiClient("{YOUR_OPENAI_API_KEY}");
string text = "Who are you?";
string response = await _client.GetChatCompletions(new UserMessage(text), maxTokens: 80);
Console.WriteLine(response);
var text = "Write the world top 3 songs of Soul genre";
await foreach (string chunk in _client.StreamChatCompletions(new UserMessage(text), maxTokens: 80))
{
Console.Write(chunk);
}
Use ThenAssistant
and ThenUser
methods to create a dialog:
var dialog = Dialog.StartAsUser("How many meters are in a kilometer? Write just the number.") //the message from user
.ThenAssistant("1000") // response from the assistant
.ThenUser("Convert it to hex. Write just the number."); // the next message from user
await foreach (var chunk in _client.StreamChatCompletions(dialog, maxTokens: 80))
{
Console.Write(chunk);
}
Or just send message history as a collection.