你当前正在访问 Microsoft Azure Global Edition 技术文档网站。 如果需要访问由世纪互联运营的 Microsoft Azure 中国技术文档网站，请访问 https://docs.azure.cn。
Azure OpenAI 支持的编程语言
适用于 .NET 的 Azure OpenAI 客户端库是适用于 .NET 的官方 OpenAI 客户端库的配套库。 Azure OpenAI 库会配置可与 Azure OpenAI 配合使用的客户端，并为特定于 Azure OpenAI 场景的请求和响应模型提供额外的强类型扩展支持。
稳定版本：
源代码 | 包 (NuGet) | 包参考文档API 参考文档 | 示例
预览版本：
预览版本有权访问最新功能。
源代码 | 包 (NuGet) | API 参考文档 | 包参考文档示例
Azure OpenAI API 版本支持
与适用于 Python 和 JavaScript 的 Azure OpenAI 客户端库不同，Azure OpenAI .NET 包仅面向 Azure OpenAI API 版本的特定子集。 通常，每个 Azure OpenAI .NET 包都会解锁对较新 Azure OpenAI API 版本功能的访问权限。 有权访问最新的 API 版本会影响功能可用性。
版本选择由 AzureOpenAIClientOptions.ServiceVersion 枚举控制。
稳定版本目前面向：
2024-06-01
预览版本目前可以面向：
2024-06-012024-08-01-preview2024-09-01-preview2024-10-01-preview
安装
dotnet add package Azure.AI.OpenAI --prerelease
Azure.AI.OpenAI 包基于官方 OpenAI 包构建，后者作为依赖项包含在内。
身份验证
若要与 Azure OpenAI 或 OpenAI 交互，请使用以下方法之一创建 AzureOpenAIClient 实例：
安全、无密钥的身份验证方法是使用 Microsoft Entra ID（前身为 Azure Active Directory）通过 Azure 标识库进行身份验证。 若要使用该库，请使用以下代码：
dotnet add package Azure.Identity
使用库中的所需凭据类型。 例如，DefaultAzureCredential：
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
ChatClient chatClient = azureClient.GetChatClient("my-gpt-4o-mini-deployment");
音频
AzureOpenAIClient.GetAudioClient
听录
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
AudioClient client = azureClient.GetAudioClient("whisper");
string audioFilePath = Path.Combine("Assets", "speech.mp3");
AudioTranscriptionOptions options = new()
{
ResponseFormat = AudioTranscriptionFormat.Verbose,
TimestampGranularities = AudioTimestampGranularities.Word | AudioTimestampGranularities.Segment,
};
AudioTranscription transcription = client.TranscribeAudio(audioFilePath, options);
Console.WriteLine("Transcription:");
Console.WriteLine($"{transcription.Text}");
Console.WriteLine();
Console.WriteLine($"Words:");
foreach (TranscribedWord word in transcription.Words)
{
Console.WriteLine($" {word.Word,15} : {word.StartTime.TotalMilliseconds,5:0} - {word.EndTime.TotalMilliseconds,5:0}");
}
Console.WriteLine();
Console.WriteLine($"Segments:");
foreach (TranscribedSegment segment in transcription.Segments)
{
Console.WriteLine($" {segment.Text,90} : {segment.StartTime.TotalMilliseconds,5:0} - {segment.EndTime.TotalMilliseconds,5:0}");
}
文本转语音 (TTS)
using Azure.AI.OpenAI;
using Azure.Identity;
using OpenAI.Audio;
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
AudioClient client = azureClient.GetAudioClient("tts-hd"); //Replace with your Azure OpenAI model deployment
string input = "Testing, testing, 1, 2, 3";
BinaryData speech = client.GenerateSpeech(input, GeneratedSpeechVoice.Alloy);
using FileStream stream = File.OpenWrite($"{Guid.NewGuid()}.mp3");
speech.ToStream().CopyTo(stream);
聊天
AzureOpenAIClient.GetChatClient
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
ChatClient chatClient = azureClient.GetChatClient("my-gpt-4o-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 recent input
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}");
流式传输聊天消息
流式聊天补全使用 CompleteChatStreaming 和 CompleteChatStreamingAsync 方法，这些方法将返回 ResultCollection<StreamingChatCompletionUpdate> 或 AsyncCollectionResult<StreamingChatCompletionUpdate>，而不是 ClientResult<ChatCompletion>。
这些结果集合可以使用 foreach 或 await foreach 来进行迭代，每次更新都是在流式响应提供新数据时到达的。
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
ChatClient chatClient = azureClient.GetChatClient("my-gpt-4o-deployment");
CollectionResult<StreamingChatCompletionUpdate> completionUpdates = chatClient.CompleteChatStreaming(
[
new SystemChatMessage("You are a helpful assistant that talks like a pirate."),
new UserChatMessage("Hi, can you help me?"),
new AssistantChatMessage("Arrr! Of course, me hearty! What can I do for ye?"),
new UserChatMessage("What's the best way to train a parrot?"),
]);
foreach (StreamingChatCompletionUpdate completionUpdate in completionUpdates)
{
foreach (ChatMessageContentPart contentPart in completionUpdate.ContentUpdate)
{
Console.Write(contentPart.Text);
}
}
嵌入
AzureOpenAIClient.GetEmbeddingClient
using Azure.AI.OpenAI;
using Azure.Identity;
using OpenAI.Embeddings;
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
EmbeddingClient client = azureClient.GetEmbeddingClient("text-embedding-3-large"); //Replace with your model deployment name
string description = "This is a test embedding";
OpenAIEmbedding embedding = client.GenerateEmbedding(description);
ReadOnlyMemory<float> vector = embedding.ToFloats();
Console.WriteLine(string.Join(", ", vector.ToArray()));
微调
目前不支持 Azure OpenAI .NET 包。
批处理
目前不支持 Azure OpenAI .NET 包。
映像
AzureOpenAIClient.GetImageClient
using Azure.AI.OpenAI;
using Azure.Identity;
using OpenAI.Images;
AzureOpenAIClient azureClient = new(
new Uri("https://your-azure-openai-resource.com"),
new DefaultAzureCredential());
ImageClient client = azureClient.GetImageClient("dall-e-3"); // replace with your model deployment name.
string prompt = "A rabbit eating pancakes.";
ImageGenerationOptions options = new()
{
Quality = GeneratedImageQuality.High,
Size = GeneratedImageSize.W1792xH1024,
Style = GeneratedImageStyle.Vivid,
ResponseFormat = GeneratedImageFormat.Bytes
};
GeneratedImage image = client.GenerateImage(prompt, options);
BinaryData bytes = image.ImageBytes;
using FileStream stream = File.OpenWrite($"{Guid.NewGuid()}.png");
bytes.ToStream().CopyTo(stream);
完成（旧版）
不支持 Azure OpenAI .NET 包。
错误处理
错误代码
| 状态代码 | 错误类型 |
|---|---|
| 400 | Bad Request Error |
| 401 | Authentication Error |
| 403 | Permission Denied Error |
| 404 | Not Found Error |
| 422 | Unprocessable Entity Error |
| 429 | Rate Limit Error |
| 500 | Internal Server Error |
| 503 | Service Unavailable |
| 504 | Gateway Timeout |
重试
客户端类将使用指数退避自动重试以下错误最多三次：
- 408 请求超时
- 429 请求次数过多
- 500 内部服务器错误
- 502 错误的网关
- 503 服务不可用
- 504 网关超时
源代码 | 包 (pkg.go.dev) | API 参考文档 | 包参考文档示例
Azure OpenAI API 版本支持
与适用于 Python 和 JavaScript 的 Azure OpenAI 客户端库不同，Azure OpenAI Go 库面向特定的 Azure OpenAI API 版本。 有权访问最新的 API 版本会影响功能可用性。
当前 Azure OpenAI API 版本目标：2024-10-01-preview
这是在 custom_client.go 文件中定义的。
安装
使用 go get 安装 azopenai 和 azidentity 模块：
go get github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai
# optional
go get github.com/Azure/azure-sdk-for-go/sdk/azidentity
身份验证
azidentity 模块用于 Azure OpenAI 的 Azure Active Directory 身份验证。
package main
import (
"log"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azidentity"
)
func main() {
dac, err := azidentity.NewDefaultAzureCredential(nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
// NOTE: this constructor creates a client that connects to an Azure OpenAI endpoint.
// To connect to the public OpenAI endpoint, use azopenai.NewClientForOpenAI
client, err := azopenai.NewClient("https://<your-azure-openai-host>.openai.azure.com", dac, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
_ = client
}
音频
Client.GenerateSpeechFromText
ackage main
import (
"context"
"fmt"
"io"
"log"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
"github.com/Azure/azure-sdk-for-go/sdk/azcore/to"
)
func main() {
openAIKey := os.Getenv("OPENAI_API_KEY")
// Ex: "https://api.openai.com/v1"
openAIEndpoint := os.Getenv("OPENAI_ENDPOINT")
modelDeploymentID := "tts-1"
if openAIKey == "" || openAIEndpoint == "" || modelDeploymentID == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(openAIKey)
client, err := azopenai.NewClientForOpenAI(openAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
audioResp, err := client.GenerateSpeechFromText(context.Background(), azopenai.SpeechGenerationOptions{
Input: to.Ptr("i am a computer"),
Voice: to.Ptr(azopenai.SpeechVoiceAlloy),
ResponseFormat: to.Ptr(azopenai.SpeechGenerationResponseFormatFlac),
DeploymentName: to.Ptr("tts-1"),
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
defer audioResp.Body.Close()
audioBytes, err := io.ReadAll(audioResp.Body)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
fmt.Fprintf(os.Stderr, "Got %d bytes of FLAC audio\n", len(audioBytes))
}
Client.GetAudioTranscription
package main
import (
"context"
"fmt"
"log"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
"github.com/Azure/azure-sdk-for-go/sdk/azcore/to"
)
func main() {
azureOpenAIKey := os.Getenv("AOAI_WHISPER_API_KEY")
// Ex: "https://<your-azure-openai-host>.openai.azure.com"
azureOpenAIEndpoint := os.Getenv("AOAI_WHISPER_ENDPOINT")
modelDeploymentID := os.Getenv("AOAI_WHISPER_MODEL")
if azureOpenAIKey == "" || azureOpenAIEndpoint == "" || modelDeploymentID == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(azureOpenAIKey)
client, err := azopenai.NewClientWithKeyCredential(azureOpenAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
mp3Bytes, err := os.ReadFile("testdata/sampledata_audiofiles_myVoiceIsMyPassportVerifyMe01.mp3")
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
resp, err := client.GetAudioTranscription(context.TODO(), azopenai.AudioTranscriptionOptions{
File: mp3Bytes,
// this will return _just_ the translated text. Other formats are available, which return
// different or additional metadata. See [azopenai.AudioTranscriptionFormat] for more examples.
ResponseFormat: to.Ptr(azopenai.AudioTranscriptionFormatText),
DeploymentName: &modelDeploymentID,
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
fmt.Fprintf(os.Stderr, "Transcribed text: %s\n", *resp.Text)
}
聊天
Client.GetChatCompletions
package main
import (
"context"
"fmt"
"log"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
)
func main() {
azureOpenAIKey := os.Getenv("AOAI_CHAT_COMPLETIONS_API_KEY")
modelDeploymentID := os.Getenv("AOAI_CHAT_COMPLETIONS_MODEL")
// Ex: "https://<your-azure-openai-host>.openai.azure.com"
azureOpenAIEndpoint := os.Getenv("AOAI_CHAT_COMPLETIONS_ENDPOINT")
if azureOpenAIKey == "" || modelDeploymentID == "" || azureOpenAIEndpoint == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(azureOpenAIKey)
// In Azure OpenAI you must deploy a model before you can use it in your client. For more information
// see here: https://learn.microsoft.com/azure/cognitive-services/openai/how-to/create-resource
client, err := azopenai.NewClientWithKeyCredential(azureOpenAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
// This is a conversation in progress.
// NOTE: all messages, regardless of role, count against token usage for this API.
messages := []azopenai.ChatRequestMessageClassification{
// You set the tone and rules of the conversation with a prompt as the system role.
&azopenai.ChatRequestSystemMessage{Content: azopenai.NewChatRequestSystemMessageContent("You are a helpful assistant. You will talk like a pirate.")},
// The user asks a question
&azopenai.ChatRequestUserMessage{Content: azopenai.NewChatRequestUserMessageContent("Can you help me?")},
// The reply would come back from the ChatGPT. You'd add it to the conversation so we can maintain context.
&azopenai.ChatRequestAssistantMessage{Content: azopenai.NewChatRequestAssistantMessageContent("Arrrr! Of course, me hearty! What can I do for ye?")},
// The user answers the question based on the latest reply.
&azopenai.ChatRequestUserMessage{Content: azopenai.NewChatRequestUserMessageContent("What's the best way to train a parrot?")},
// from here you'd keep iterating, sending responses back from ChatGPT
}
gotReply := false
resp, err := client.GetChatCompletions(context.TODO(), azopenai.ChatCompletionsOptions{
// This is a conversation in progress.
// NOTE: all messages count against token usage for this API.
Messages: messages,
DeploymentName: &modelDeploymentID,
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
for _, choice := range resp.Choices {
gotReply = true
if choice.ContentFilterResults != nil {
fmt.Fprintf(os.Stderr, "Content filter results\n")
if choice.ContentFilterResults.Error != nil {
fmt.Fprintf(os.Stderr, " Error:%v\n", choice.ContentFilterResults.Error)
}
fmt.Fprintf(os.Stderr, " Hate: sev: %v, filtered: %v\n", *choice.ContentFilterResults.Hate.Severity, *choice.ContentFilterResults.Hate.Filtered)
fmt.Fprintf(os.Stderr, " SelfHarm: sev: %v, filtered: %v\n", *choice.ContentFilterResults.SelfHarm.Severity, *choice.ContentFilterResults.SelfHarm.Filtered)
fmt.Fprintf(os.Stderr, " Sexual: sev: %v, filtered: %v\n", *choice.ContentFilterResults.Sexual.Severity, *choice.ContentFilterResults.Sexual.Filtered)
fmt.Fprintf(os.Stderr, " Violence: sev: %v, filtered: %v\n", *choice.ContentFilterResults.Violence.Severity, *choice.ContentFilterResults.Violence.Filtered)
}
if choice.Message != nil && choice.Message.Content != nil {
fmt.Fprintf(os.Stderr, "Content[%d]: %s\n", *choice.Index, *choice.Message.Content)
}
if choice.FinishReason != nil {
// this choice's conversation is complete.
fmt.Fprintf(os.Stderr, "Finish reason[%d]: %s\n", *choice.Index, *choice.FinishReason)
}
}
if gotReply {
fmt.Fprintf(os.Stderr, "Got chat completions reply\n")
}
}
Client.GetChatCompletionsStream
package main
import (
"context"
"errors"
"fmt"
"io"
"log"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
"github.com/Azure/azure-sdk-for-go/sdk/azcore/to"
)
func main() {
azureOpenAIKey := os.Getenv("AOAI_CHAT_COMPLETIONS_API_KEY")
modelDeploymentID := os.Getenv("AOAI_CHAT_COMPLETIONS_MODEL")
// Ex: "https://<your-azure-openai-host>.openai.azure.com"
azureOpenAIEndpoint := os.Getenv("AOAI_CHAT_COMPLETIONS_ENDPOINT")
if azureOpenAIKey == "" || modelDeploymentID == "" || azureOpenAIEndpoint == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(azureOpenAIKey)
// In Azure OpenAI you must deploy a model before you can use it in your client. For more information
// see here: https://learn.microsoft.com/azure/cognitive-services/openai/how-to/create-resource
client, err := azopenai.NewClientWithKeyCredential(azureOpenAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
// This is a conversation in progress.
// NOTE: all messages, regardless of role, count against token usage for this API.
messages := []azopenai.ChatRequestMessageClassification{
// You set the tone and rules of the conversation with a prompt as the system role.
&azopenai.ChatRequestSystemMessage{Content: azopenai.NewChatRequestSystemMessageContent("You are a helpful assistant. You will talk like a pirate and limit your responses to 20 words or less.")},
// The user asks a question
&azopenai.ChatRequestUserMessage{Content: azopenai.NewChatRequestUserMessageContent("Can you help me?")},
// The reply would come back from the ChatGPT. You'd add it to the conversation so we can maintain context.
&azopenai.ChatRequestAssistantMessage{Content: azopenai.NewChatRequestAssistantMessageContent("Arrrr! Of course, me hearty! What can I do for ye?")},
// The user answers the question based on the latest reply.
&azopenai.ChatRequestUserMessage{Content: azopenai.NewChatRequestUserMessageContent("What's the best way to train a parrot?")},
// from here you'd keep iterating, sending responses back from ChatGPT
}
resp, err := client.GetChatCompletionsStream(context.TODO(), azopenai.ChatCompletionsStreamOptions{
// This is a conversation in progress.
// NOTE: all messages count against token usage for this API.
Messages: messages,
N: to.Ptr[int32](1),
DeploymentName: &modelDeploymentID,
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
defer resp.ChatCompletionsStream.Close()
gotReply := false
for {
chatCompletions, err := resp.ChatCompletionsStream.Read()
if errors.Is(err, io.EOF) {
break
}
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
for _, choice := range chatCompletions.Choices {
gotReply = true
text := ""
if choice.Delta.Content != nil {
text = *choice.Delta.Content
}
role := ""
if choice.Delta.Role != nil {
role = string(*choice.Delta.Role)
}
fmt.Fprintf(os.Stderr, "Content[%d], role %q: %q\n", *choice.Index, role, text)
}
}
if gotReply {
fmt.Fprintf(os.Stderr, "Got chat completions streaming reply\n")
}
}
嵌入
Client.GetEmbeddings
package main
import (
"context"
"fmt"
"log"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
)
func main() {
azureOpenAIKey := os.Getenv("AOAI_EMBEDDINGS_API_KEY")
modelDeploymentID := os.Getenv("AOAI_EMBEDDINGS_MODEL")
// Ex: "https://<your-azure-openai-host>.openai.azure.com"
azureOpenAIEndpoint := os.Getenv("AOAI_EMBEDDINGS_ENDPOINT")
if azureOpenAIKey == "" || modelDeploymentID == "" || azureOpenAIEndpoint == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(azureOpenAIKey)
// In Azure OpenAI you must deploy a model before you can use it in your client. For more information
// see here: https://learn.microsoft.com/azure/cognitive-services/openai/how-to/create-resource
client, err := azopenai.NewClientWithKeyCredential(azureOpenAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
resp, err := client.GetEmbeddings(context.TODO(), azopenai.EmbeddingsOptions{
Input: []string{"Testing, testing, 1,2,3."},
DeploymentName: &modelDeploymentID,
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
for _, embed := range resp.Data {
// embed.Embedding contains the embeddings for this input index.
fmt.Fprintf(os.Stderr, "Got embeddings for input %d\n", *embed.Index)
}
}
图像生成
Client.GetImageGenerations
package main
import (
"context"
"fmt"
"log"
"net/http"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
"github.com/Azure/azure-sdk-for-go/sdk/azcore/to"
)
func main() {
azureOpenAIKey := os.Getenv("AOAI_DALLE_API_KEY")
// Ex: "https://<your-azure-openai-host>.openai.azure.com"
azureOpenAIEndpoint := os.Getenv("AOAI_DALLE_ENDPOINT")
azureDeployment := os.Getenv("AOAI_DALLE_MODEL")
if azureOpenAIKey == "" || azureOpenAIEndpoint == "" || azureDeployment == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(azureOpenAIKey)
client, err := azopenai.NewClientWithKeyCredential(azureOpenAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
resp, err := client.GetImageGenerations(context.TODO(), azopenai.ImageGenerationOptions{
Prompt: to.Ptr("a cat"),
ResponseFormat: to.Ptr(azopenai.ImageGenerationResponseFormatURL),
DeploymentName: &azureDeployment,
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
for _, generatedImage := range resp.Data {
// the underlying type for the generatedImage is dictated by the value of
// ImageGenerationOptions.ResponseFormat. In this example we used `azopenai.ImageGenerationResponseFormatURL`,
// so the underlying type will be ImageLocation.
resp, err := http.Head(*generatedImage.URL)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
_ = resp.Body.Close()
fmt.Fprintf(os.Stderr, "Image generated, HEAD request on URL returned %d\n", resp.StatusCode)
}
}
完成（旧版）
Client.GetChatCompletions
package main
import (
"context"
"fmt"
"log"
"os"
"github.com/Azure/azure-sdk-for-go/sdk/ai/azopenai"
"github.com/Azure/azure-sdk-for-go/sdk/azcore"
"github.com/Azure/azure-sdk-for-go/sdk/azcore/to"
)
func main() {
azureOpenAIKey := os.Getenv("AOAI_COMPLETIONS_API_KEY")
modelDeployment := os.Getenv("AOAI_COMPLETIONS_MODEL")
// Ex: "https://<your-azure-openai-host>.openai.azure.com"
azureOpenAIEndpoint := os.Getenv("AOAI_COMPLETIONS_ENDPOINT")
if azureOpenAIKey == "" || modelDeployment == "" || azureOpenAIEndpoint == "" {
fmt.Fprintf(os.Stderr, "Skipping example, environment variables missing\n")
return
}
keyCredential := azcore.NewKeyCredential(azureOpenAIKey)
// In Azure OpenAI you must deploy a model before you can use it in your client. For more information
// see here: https://learn.microsoft.com/azure/cognitive-services/openai/how-to/create-resource
client, err := azopenai.NewClientWithKeyCredential(azureOpenAIEndpoint, keyCredential, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
resp, err := client.GetCompletions(context.TODO(), azopenai.CompletionsOptions{
Prompt: []string{"What is Azure OpenAI, in 20 words or less"},
MaxTokens: to.Ptr(int32(2048)),
Temperature: to.Ptr(float32(0.0)),
DeploymentName: &modelDeployment,
}, nil)
if err != nil {
// TODO: Update the following line with your application specific error handling logic
log.Printf("ERROR: %s", err)
return
}
for _, choice := range resp.Choices {
fmt.Fprintf(os.Stderr, "Result: %s\n", *choice.Text)
}
}
错误处理
发送 HTTP 请求的所有方法都会在这些请求失败时返回 *azcore.ResponseError。
ResponseError 包含错误详细信息和来自服务的原始响应。
Logging
此模块使用 azcore 中的日志记录实现。 若要为所有 Azure SDK 模块启用日志记录，请将 AZURE_SDK_GO_LOGGING 设置为 all。 默认情况下，记录器将写入 stderr。 请使用 azcore/log 包控制日志输出。 例如，仅在日志中记录 HTTP 请求和响应事件，并将其打印到 stdout：
import azlog "github.com/Azure/azure-sdk-for-go/sdk/azcore/log"
// Print log events to stdout
azlog.SetListener(func(cls azlog.Event, msg string) {
fmt.Println(msg)
})
// Includes only requests and responses in credential logs
azlog.SetEvents(azlog.EventRequest, azlog.EventResponse)
源代码 | 工件 (Maven) | API 参考文档 | 包参考文档示例
Azure OpenAI API 版本支持
与适用于 Python 和 JavaScript 的 Azure OpenAI 客户端库不同，为了确保兼容性，Azure OpenAI Java 包仅面向 Azure OpenAI API 版本的特定子集。 通常，每个 Azure OpenAI Java 包都会解锁对较新 Azure OpenAI API 版本功能的访问权限。 有权访问最新的 API 版本会影响功能可用性。
版本选择由 OpenAIServiceVersion 枚举控制。
支持的最新 Azure OpenAI 预览版 API 为：
-2024-08-01-preview
支持的最新稳定版（正式发布）为：
-2024-06-01
安装
包详细信息
<dependency>
<groupId>com.azure</groupId>
<artifactId>azure-ai-openai</artifactId>
<version>1.0.0-beta.12</version>
</dependency>
身份验证
若要与 Azure OpenAI 服务交互，你需要使用 OpenAIClientBuilder 创建客户端类 OpenAIAsyncClient 或 OpenAIClient 的实例。 若要配置可与 Azure OpenAI 配合使用的客户端，请提供 Azure OpenAI 资源的有效终结点 URI 以及有权使用 Azure OpenAI 资源的相应密钥凭据、令牌凭据或 Azure 标识凭据。
使用 Microsoft Entra ID 进行身份验证需要一些初始设置：
添加 Azure 标识包：
<dependency>
<groupId>com.azure</groupId>
<artifactId>azure-identity</artifactId>
<version>1.13.3</version>
</dependency>
设置后，可以从 azure.identity 中选择要使用的凭据类型。 例如，DefaultAzureCredential 可用于对客户端进行身份验证：将 Microsoft Entra ID 应用程序的客户端 ID、租户 ID 和客户端密码的值设置为环境变量：AZURE_CLIENT_ID、AZURE_TENANT_ID、AZURE_CLIENT_SECRET。
使用 DefaultAzureCredential 时授权最简单。 它会在其运行环境中查找要使用的最佳凭据。
TokenCredential defaultCredential = new DefaultAzureCredentialBuilder().build();
OpenAIClient client = new OpenAIClientBuilder()
.credential(defaultCredential)
.endpoint("{endpoint}")
.buildClient();
音频
client.getAudioTranscription
String fileName = "{your-file-name}";
Path filePath = Paths.get("{your-file-path}" + fileName);
byte[] file = BinaryData.fromFile(filePath).toBytes();
AudioTranscriptionOptions transcriptionOptions = new AudioTranscriptionOptions(file)
.setResponseFormat(AudioTranscriptionFormat.JSON);
AudioTranscription transcription = client.getAudioTranscription("{deploymentOrModelName}", fileName, transcriptionOptions);
System.out.println("Transcription: " + transcription.getText());
client.generateSpeechFromText
文本转语音 (TTS)
String deploymentOrModelId = "{azure-open-ai-deployment-model-id}";
SpeechGenerationOptions options = new SpeechGenerationOptions(
"Today is a wonderful day to build something people love!",
SpeechVoice.ALLOY);
BinaryData speech = client.generateSpeechFromText(deploymentOrModelId, options);
// Checkout your generated speech in the file system.
Path path = Paths.get("{your-local-file-path}/speech.wav");
Files.write(path, speech.toBytes());
聊天
client.getChatCompletions
List<ChatRequestMessage> chatMessages = new ArrayList<>();
chatMessages.add(new ChatRequestSystemMessage("You are a helpful assistant. You will talk like a pirate."));
chatMessages.add(new ChatRequestUserMessage("Can you help me?"));
chatMessages.add(new ChatRequestAssistantMessage("Of course, me hearty! What can I do for ye?"));
chatMessages.add(new ChatRequestUserMessage("What's the best way to train a parrot?"));
ChatCompletions chatCompletions = client.getChatCompletions("{deploymentOrModelName}",
new ChatCompletionsOptions(chatMessages));
System.out.printf("Model ID=%s is created at %s.%n", chatCompletions.getId(), chatCompletions.getCreatedAt());
for (ChatChoice choice : chatCompletions.getChoices()) {
ChatResponseMessage message = choice.getMessage();
System.out.printf("Index: %d, Chat Role: %s.%n", choice.getIndex(), message.getRole());
System.out.println("Message:");
System.out.println(message.getContent());
}
流式处理
List<ChatRequestMessage> chatMessages = new ArrayList<>();
chatMessages.add(new ChatRequestSystemMessage("You are a helpful assistant. You will talk like a pirate."));
chatMessages.add(new ChatRequestUserMessage("Can you help me?"));
chatMessages.add(new ChatRequestAssistantMessage("Of course, me hearty! What can I do for ye?"));
chatMessages.add(new ChatRequestUserMessage("What's the best way to train a parrot?"));
ChatCompletions chatCompletions = client.getChatCompletions("{deploymentOrModelName}",
new ChatCompletionsOptions(chatMessages));
System.out.printf("Model ID=%s is created at %s.%n", chatCompletions.getId(), chatCompletions.getCreatedAt());
for (ChatChoice choice : chatCompletions.getChoices()) {
ChatResponseMessage message = choice.getMessage();
System.out.printf("Index: %d, Chat Role: %s.%n", choice.getIndex(), message.getRole());
System.out.println("Message:");
System.out.println(message.getContent());
}
关于图片的聊天补全
List<ChatRequestMessage> chatMessages = new ArrayList<>();
chatMessages.add(new ChatRequestSystemMessage("You are a helpful assistant that describes images"));
chatMessages.add(new ChatRequestUserMessage(Arrays.asList(
new ChatMessageTextContentItem("Please describe this image"),
new ChatMessageImageContentItem(
new ChatMessageImageUrl("https://raw.githubusercontent.com/MicrosoftDocs/azure-ai-docs/main/articles/ai-services/openai/media/how-to/generated-seattle.png"))
)));
ChatCompletionsOptions chatCompletionsOptions = new ChatCompletionsOptions(chatMessages);
ChatCompletions chatCompletions = client.getChatCompletions("{deploymentOrModelName}", chatCompletionsOptions);
System.out.println("Chat completion: " + chatCompletions.getChoices().get(0).getMessage().getContent());
嵌入
client.getEmbeddings
EmbeddingsOptions embeddingsOptions = new EmbeddingsOptions(
Arrays.asList("Your text string goes here"));
Embeddings embeddings = client.getEmbeddings("{deploymentOrModelName}", embeddingsOptions);
for (EmbeddingItem item : embeddings.getData()) {
System.out.printf("Index: %d.%n", item.getPromptIndex());
for (Float embedding : item.getEmbedding()) {
System.out.printf("%f;", embedding);
}
}
图像生成
ImageGenerationOptions imageGenerationOptions = new ImageGenerationOptions(
"A drawing of the Seattle skyline in the style of Van Gogh");
ImageGenerations images = client.getImageGenerations("{deploymentOrModelName}", imageGenerationOptions);
for (ImageGenerationData imageGenerationData : images.getData()) {
System.out.printf(
"Image location URL that provides temporary access to download the generated image is %s.%n",
imageGenerationData.getUrl());
}
处理错误
启用客户端日志记录
若要排查 Azure OpenAI 库的问题，请务必首先启用日志记录功能来监视应用程序的行为。 日志中的错误和警告通常会提供有关所出现问题的有用见解，有时还会包括可用于修复问题的纠正措施。 适用于 Java 的 Azure 客户端库具有两个日志记录选项：
- 内置日志记录框架。
- 支持使用 SLF4J 接口进行日志记录。
请参阅参考文档中的说明，了解如何[在 Azure SDK for Java 中配置日志记录功能][logging_overview]。
启用 HTTP 请求/响应日志记录
查看通过网络发送到 Azure OpenAI 服务或接收自该服务的 HTTP 请求或响应在排查问题时非常有用。 若要启用 HTTP 请求和响应有效负载的日志记录，可以按如下所示配置 [OpenAIClient][openai_client]。 如果类路径上没有 SLF4J 的 Logger，请在计算机中设置环境变量 [AZURE_LOG_LEVEL][azure_log_level] 以启用日志记录。
OpenAIClient openAIClient = new OpenAIClientBuilder()
.endpoint("{endpoint}")
.credential(new AzureKeyCredential("{key}"))
.httpLogOptions(new HttpLogOptions().setLogLevel(HttpLogDetailLevel.BODY_AND_HEADERS))
.buildClient();
// or
DefaultAzureCredential credential = new DefaultAzureCredentialBuilder().build();
OpenAIClient configurationClientAad = new OpenAIClientBuilder()
.credential(credential)
.endpoint("{endpoint}")
.httpLogOptions(new HttpLogOptions().setLogLevel(HttpLogDetailLevel.BODY_AND_HEADERS))
.buildClient();
或者，可以通过设置以下环境变量来为整个应用程序配置 HTTP 请求和响应的日志记录。 请注意，此更改将为支持 HTTP 请求/响应日志记录的每个 Azure 客户端启用日志记录。
环境变量名称：AZURE_HTTP_LOG_DETAIL_LEVEL
| 值 | 日志记录级别 |
|---|---|
| 无 | HTTP 请求/响应日志记录已禁用 |
| 基本 | 仅记录 URL、HTTP 方法和完成请求的时间。 |
| headers | 记录 BASIC 中的所有内容，加上所有请求和响应头。 |
| 正文 | 记录 BASIC 中的所有内容，加上所有请求和响应主体。 |
| body_and_headers | 记录 HEADERS 和 BODY 中的所有内容。 |
注意
记录请求和响应的正文时，请确保其中不包含机密信息。 记录标头时，客户端库有一组默认标头被视为可以安全记录，但可以通过更新生成器中的日志选项来更新这组标头，如下所示。
clientBuilder.httpLogOptions(new HttpLogOptions().addAllowedHeaderName("safe-to-log-header-name"))
针对异常的故障排除
Azure OpenAI 服务方法会在失败时引发 [HttpResponseException 或其子类。
OpenAI 客户端库引发的 HttpResponseException 包括详细的响应错误对象，该对象提供有关出错情况的特定有用见解，并包括可用于修复常见问题的纠正措施。
你可以在 HttpResponseException 对象的消息属性中找到此错误信息。
下面是如何使用同步客户端捕获它的示例
List<ChatRequestMessage> chatMessages = new ArrayList<>();
chatMessages.add(new ChatRequestSystemMessage("You are a helpful assistant. You will talk like a pirate."));
chatMessages.add(new ChatRequestUserMessage("Can you help me?"));
chatMessages.add(new ChatRequestAssistantMessage("Of course, me hearty! What can I do for ye?"));
chatMessages.add(new ChatRequestUserMessage("What's the best way to train a parrot?"));
try {
ChatCompletions chatCompletions = client.getChatCompletions("{deploymentOrModelName}",
new ChatCompletionsOptions(chatMessages));
} catch (HttpResponseException e) {
System.out.println(e.getMessage());
// Do something with the exception
}
使用异步客户端，可以捕获和处理错误回调中的异常：
asyncClient.getChatCompletions("{deploymentOrModelName}", new ChatCompletionsOptions(chatMessages))
.doOnSuccess(ignored -> System.out.println("Success!"))
.doOnError(
error -> error instanceof ResourceNotFoundException,
error -> System.out.println("Exception: 'getChatCompletions' could not be performed."));
身份验证错误
Azure OpenAI 支持 Microsoft Entra ID 身份验证。
OpenAIClientBuilder 具有设置 credential 的方法。 若要提供有效的凭据，可以使用 azure-identity 依赖项。
Azure OpenAI API 版本支持
Azure OpenAI 中的功能可用性取决于你面向的 REST API 版本。 如需使用最新功能，请面向最新的预览版 API。
| 最新正式版 API | 最新预览版 API |
|---|---|
2024-10-21 |
2025-01-01-preview |
安装
npm install openai
身份验证
可通过若干方式使用 Microsoft Entra ID 令牌向 Azure OpenAI 服务进行身份验证。 默认方法是使用 @azure/identity 包中的 DefaultAzureCredential 类。
import { DefaultAzureCredential } from "@azure/identity";
const credential = new DefaultAzureCredential();
然后，该对象会传递给 OpenAIClient 和 AssistantsClient 客户端构造函数的第二个参数。
但是，为了对 AzureOpenAI 客户端进行身份验证，我们需要使用 @azure/identity 包中的 getBearerTokenProvider 函数。 该函数创建一个令牌提供程序，AzureOpenAI 内部使用该提供程序为每个请求获取令牌。 令牌提供程序的创建方式如下：
import { AzureOpenAI } from 'openai';
import { DefaultAzureCredential, getBearerTokenProvider } from "@azure/identity";
const credential = new DefaultAzureCredential();
const endpoint = "https://your-azure-openai-resource.com";
const apiVersion = "2024-10-21"
const scope = "https://cognitiveservices.azure.com/.default";
const azureADTokenProvider = getBearerTokenProvider(credential, scope);
const client = new AzureOpenAI({
endpoint,
apiVersions,
azureADTokenProvider
});
音频
听录
import { createReadStream } from "fs";
const result = await client.audio.transcriptions.create({
model: '',
file: createReadStream(audioFilePath),
});
聊天
chat.completions.create
const result = await client.chat.completions.create({ messages, model: '', max_tokens: 100 });
流式处理
const stream = await client.chat.completions.create({ model: '', messages, max_tokens: 100, stream: true });
嵌入
const embeddings = await client.embeddings.create({ input, model: '' });
图像生成
const results = await client.images.generate({ prompt, model: '', n, size });
错误处理
错误代码
| 状态代码 | 错误类型 |
|---|---|
| 400 | Bad Request Error |
| 401 | Authentication Error |
| 403 | Permission Denied Error |
| 404 | Not Found Error |
| 422 | Unprocessable Entity Error |
| 429 | Rate Limit Error |
| 500 | Internal Server Error |
| 503 | Service Unavailable |
| 504 | Gateway Timeout |
重试
默认情况下，以下错误会自动停用两次，并出现短暂的指数退避：
- 连接错误
- 408 请求超时
- 429 速率限制
-
>=500 内部错误
使用 maxRetries 设置/禁用重试行为：
// Configure the default for all requests:
const client = new AzureOpenAI({
maxRetries: 0, // default is 2
});
// Or, configure per-request:
await client.chat.completions.create({ messages: [{ role: 'user', content: 'How can I get the name of the current day in Node.js?' }], model: '' }, {
maxRetries: 5,
});
注意
此库由 OpenAI 维护。 请参阅发布历史记录，以跟踪库的最新更新。
Azure OpenAI API 版本支持
Azure OpenAI 中的功能可用性取决于你面向的 REST API 版本。 如需使用最新功能，请面向最新的预览版 API。
| 最新正式版 API | 最新预览版 API |
|---|---|
2024-10-21 |
2025-01-01-preview |
安装
pip install openai
有关最新版本：
pip install openai --upgrade
身份验证
import os
from openai import AzureOpenAI
from azure.identity import DefaultAzureCredential, get_bearer_token_provider
token_provider = get_bearer_token_provider(
DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default"
)
client = AzureOpenAI(
azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT"),
azure_ad_token_provider=token_provider,
api_version="2024-10-21"
)
音频
audio.speech.create()
此函数当前需要 API 预览版。
设置 api_version="2024-10-01-preview" 以使用此函数。
# from openai import AzureOpenAI
# client = AzureOpenAI()
from pathlib import Path
import os
speech_file_path = Path("speech.mp3")
response = client.audio.speech.create(
model="tts-hd", #Replace with model deployment name
voice="alloy",
input="Testing, testing, 1,2,3."
)
response.write_to_file(speech_file_path)
audio.transcriptions.create()
# from openai import AzureOpenAI
# client = AzureOpenAI()
audio_file = open("speech1.mp3", "rb")
transcript = client.audio.transcriptions.create(
model="whisper", # Replace with model deployment name
file=audio_file
)
print(transcript)
聊天
chat.completions.create()
# from openai import AzureOpenAI
# client = AzureOpenAI()
completion = client.chat.completions.create(
model="gpt-4o", # Replace with your model dpeloyment name.
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "When was Microsoft founded?"}
]
)
#print(completion.choices[0].message)
print(completion.model_dump_json(indent=2)
chat.completions.create() - 流式处理
# from openai import AzureOpenAI
# client = AzureOpenAI()
completion = client.chat.completions.create(
model="gpt-4o", # Replace with your model dpeloyment name.
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "When was Microsoft founded?"}
],
stream=True
)
for chunk in completion:
if chunk.choices and chunk.choices[0].delta.content is not None:
print(chunk.choices[0].delta.content, end='',)
chat.completions.create() - 图像输入
completion = client.chat.completions.create(
model="gpt-4o",
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "What's in this image?"},
{
"type": "image_url",
"image_url": {
"url": "https://raw.githubusercontent.com/MicrosoftDocs/azure-ai-docs/main/articles/ai-services/openai/media/how-to/generated-seattle.png",
}
},
],
}
],
max_tokens=300,
)
print(completion.model_dump_json(indent=2))
嵌入
embeddings.create()
# from openai import AzureOpenAI
# client = AzureOpenAI()
embedding = client.embeddings.create(
model="text-embedding-3-large", # Replace with your model deployment name
input="Attenion is all you need",
encoding_format="float"
)
print(embedding)
微调
批处理
映像
images.generate()
# from openai import AzureOpenAI
# client = AzureOpenAI()
generate_image = client.images.generate(
model="dall-e-3", #replace with your model deployment name
prompt="A rabbit eating pancakes",
n=1,
size="1024x1024",
quality = "hd",
response_format = "url",
style = "vivid"
)
print(generate_image.model_dump_json(indent=2))
完成（旧版）
completions.create()
# from openai import AzureOpenAI
# client = AzureOpenAI()
legacy_completion = client.completions.create(
model="gpt-35-turbo-instruct", # Replace with model deployment name
prompt="Hello World!",
max_tokens=100,
temperature=0
)
print(legacy_completion.model_dump_json(indent=2))
错误处理
# from openai import AzureOpenAI
# client = AzureOpenAI()
import openai
try:
client.fine_tuning.jobs.create(
model="gpt-4o",
training_file="file-test",
)
except openai.APIConnectionError as e:
print("The server could not be reached")
print(e.__cause__) # an underlying Exception, likely raised within httpx.
except openai.RateLimitError as e:
print("A 429 status code was received; we should back off a bit.")
except openai.APIStatusError as e:
print("Another non-200-range status code was received")
print(e.status_code)
print(e.response)
错误代码
| 状态代码 | 错误类型 |
|---|---|
| 400 | BadRequestError |
| 401 | AuthenticationError |
| 403 | PermissionDeniedError |
| 404 | NotFoundError |
| 422 | UnprocessableEntityError |
| 429 | RateLimitError |
| >=500 | InternalServerError |
| 空值 | APIConnectionError |
请求 ID
若要检索请求的 ID，可以使用对应于 x-request-id 响应头的 _request_id 属性。
print(completion._request_id)
print(legacy_completion._request_id)
重试
默认情况下，以下错误会自动停用两次，并出现短暂的指数退避：
- 连接错误
- 408 请求超时
- 429 速率限制
-
>=500 内部错误
使用 max_retries 设置/禁用重试行为：
# For all requests
from openai import AzureOpenAI
client = AzureOpenAI(
max_retries=0
)
# max retires for specific requests
client.with_options(max_retries=5).chat.completions.create(
messages=[
{
"role": "user",
"content": "When was Microsoft founded?",
}
],
model="gpt-4o",
)
后续步骤
- 若要查看当前支持哪些模型，请查看 Azure OpenAI 模型页