Skip to content

Latest commit

 

History

History
240 lines (178 loc) · 16.2 KB

CHANGELOG.md

File metadata and controls

240 lines (178 loc) · 16.2 KB
Raw

Release History

1.0.0-beta.10 (2023-12-06)

Following OpenAI's November Dev Day and Microsoft's 2023 Ignite conference, this update brings a slew of new features and changes to the SDK.

Features Added

  • -1106 model feature support for gpt-35-turbo and gpt-4-turbo, including use of seed, system_fingerprint, parallel function calling via tools, "JSON mode" for guaranteed function outputs, and more
  • dall-e-3 image generation capabilities via GetImageGenerations, featuring higher model quality, automatic prompt revisions by gpt-4, and customizable quality/style settings
  • Greatly expanded "On Your Data" capabilities in Azure OpenAI, including many new data source options and authentication mechanisms
  • Early support for gpt-4-vision-preview, which allows the hybrid use of text and images as input to enable scenarios like "describe this image for me"
  • Support for Azure enhancements to gpt-4-vision-preview results that include grounding and OCR features

Breaking Changes

ChatMessage changes:

  • The singular ChatMessage type has been replaced by ChatRequestMessage and ChatResponseMessage, the former of which is an abstract, polymorphic type with concrete derivations like ChatRequestSystemMessage and ChatRequestUserMessage. This requires conversion from old ChatMessages into the new types. While this is usually a straightforward string replacement, converting a response message into a request message (e.g. when propagating an assistant response to continue the conversation) will require creating a new instance of the appropriate request message with the response message's data. See the examples for details.

Dall-e-3:

  • Azure OpenAI now uses dall-e-3 model deployments for its image generation API and such a valid deployment must be provided into the options for the GetImageGenerations method to receive results.

Other changes

  • Audio transcription and translation (via GetAudioTranscription() and GetAudioTranslation() now allow specification of an optional Filename in addition to the binary audio data. This is used purely as an identifier and does not functionally alter the transcription/translation behavior in any way.

1.0.0-beta.9 (2023-11-06)

Breaking Changes

This update includes a number of version-to-version breaking changes to the API.

Streaming for completions and chat completions

Streaming Completions and Streaming Chat Completions have been significantly updated to use simpler, shallower usage patterns and data representations. The goal of these changes is to make streaming much easier to consume in common cases while still retaining full functionality in more complex ones (e.g. with multiple choices requested).

  • A new StreamingResponse<T> type is introduced that implicitly exposes an IAsyncEnumerable<T> derived from the underlying response.
  • OpenAI.GetCompletionsStreaming() now returns a StreamingResponse<Completions> that may be directly enumerated over. StreamingCompletions, StreamingChoice, and the corresponding methods are removed.
  • Because Chat Completions use a distinct structure for their streaming response messages, a new StreamingChatCompletionsUpdate type is introduced that encapsulates this update data.
  • Correspondingly, OpenAI.GetChatCompletionsStreaming() now returns a StreamingResponse<StreamingChatCompletionsUpdate> that may be enumerated over directly. StreamingChatCompletions, StreamingChatChoice, and related methods are removed.
  • For more information, please see the related pull request description as well as the updated snippets in the project README.

deploymentOrModelName moved to *Options.DeploymentName

deploymentOrModelName and related method parameters on OpenAIClient have been moved to DeploymentName properties in the corresponding method options. This is intended to promote consistency across scenario, language, and Azure/non-Azure OpenAI use.

As an example, the following:

ChatCompletionsOptions chatCompletionsOptions = new()
{
    Messages = { new(ChatRole.User, "Hello, assistant!") },
};
Response<ChatCompletions> response = client.GetChatCompletions("gpt-4", chatCompletionsOptions);

...is now re-written as:

ChatCompletionsOptions chatCompletionsOptions = new()
{
    DeploymentName = "gpt-4",
    Messages = { new(ChatRole.User, "Hello, assistant!") },
};
Response<ChatCompletions> response = client.GetChatCompletions(chatCompletionsOptions);

Consistency in complex method options type constructors

With the migration of DeploymentName into method complex options types, these options types have now been snapped to follow a common pattern: each complex options type will feature a default constructor that allows init-style setting of properties as well as a single additional constructor that accepts all required parameters for the corresponding method. Existing constructors that no longer meet that "all" requirement, including those impacted by the addition of DeploymentName, have been removed. The "convenience" constructors that represented required parameter data differently -- for example, EmbeddingsOptions(string), have also been removed in favor of the consistent "set of directly provide" choice.

More exhaustively, removed are:

  • AudioTranscriptionOptions(BinaryData)
  • AudioTranslationOptions(BinaryData)
  • ChatCompletionsOptions(IEnumerable<ChatMessage>)
  • CompletionsOptions(IEnumerable<string>)
  • EmbeddingsOptions(string)
  • EmbeddingsOptions(IEnumerable<string>)

And added as replacements are:

  • AudioTranscriptionOptions(string, BinaryData)
  • AudioTranslationOptions(string, BinaryData)
  • ChatCompletionsOptions(string, IEnumerable<ChatMessage>)
  • CompletionsOptions(string, IEnumerable<string>)
  • EmbeddingsOptions(string, IEnumerable<string>)

Embeddings now represented as ReadOnlyMemory<float>

Changed the representation of embeddings (specifically, the type of the Embedding property of the EmbeddingItem class) from IReadOnlyList<float> to ReadOnlyMemory<float> as part of a broader effort to establish consistency across the .NET ecosystem.

SearchKey and EmbeddingKey properties replaced by SetSearchKey and SetEmbeddingKey methods

Replaced the SearchKey and EmbeddingKey properties of the AzureCognitiveSearchChatExtensionConfiguration class with new SetSearchKey and SetEmbeddingKey methods respectively. These methods simplify the configuration of the Azure Cognitive Search chat extension by receiving a plain string instead of an AzureKeyCredential, promote more sensible key and secret management, and align with the Azure SDK guidelines.

1.0.0-beta.8 (2023-09-21)

Features Added

  • Audio Transcription and Audio Translation using OpenAI Whisper models is now supported. See OpenAI's API reference or the Azure OpenAI quickstart for detailed overview and background information.
    • The new methods GetAudioTranscription and GetAudioTranscription expose these capabilities on OpenAIClient
    • Transcription produces text in the primary, supported, spoken input language of the audio data provided, together with any optional associated metadata
    • Translation produces text, translated to English and reflective of the audio data provided, together with any optional associated metadata
    • These methods work for both Azure OpenAI and non-Azure api.openai.com client configurations

Breaking Changes

  • The underlying representation of PromptFilterResults (for Completions and ChatCompletions) has had its response body key changed from prompt_annotations to prompt_filter_results
  • Prior versions of the Azure.AI.OpenAI library may no longer populate PromptFilterResults as expected and it's highly recommended to upgrade to this version if the use of Azure OpenAI content moderation annotations for input data is desired
  • If a library version upgrade is not immediately possible, it's advised to use Response<T>.GetRawResponse() and manually extract the prompt_filter_results object from the top level of the Completions or ChatCompletions response Content payload

Bugs Fixed

  • Support for the described breaking change for PromptFilterResults was added and this library version will now again deserialize PromptFilterResults appropriately
  • PromptFilterResults and ContentFilterResults are now exposed on the result classes for streaming Completions and Chat Completions. Streaming(Chat)Completions.PromptFilterResults will report an index-sorted list of all prompt annotations received so far while Streaming(Chat)Choice.ContentFilterResults will reflect the latest-received content annotations that were populated and received while streaming

1.0.0-beta.7 (2023-08-25)

Features Added

  • The Azure OpenAI "using your own data" feature is now supported. See the Azure OpenAI using your own data quickstart for conceptual background and detailed setup instructions.
    • Azure OpenAI chat extensions are configured via a new AzureChatExtensionsOptions property on ChatCompletionsOptions. When an AzureChatExtensionsOptions is provided, configured requests will only work with clients configured to use the Azure OpenAI service, as the capabilities are unique to that service target.
    • AzureChatExtensionsOptions then has AzureChatExtensionConfiguration instances added to its Extensions property, with these instances representing the supplementary information needed for Azure OpenAI to use desired data sources to supplement chat completions behavior.
    • ChatChoice instances on a ChatCompletions response value that used chat extensions will then also have their Message property supplemented by an AzureChatExtensionMessageContext instance. This context contains a collection of supplementary Messages that describe the behavior of extensions that were used and supplementary response data, such as citations, provided along with the response.
    • See the README sample snippet for a simplified example of request/response use with "using your own data"

1.0.0-beta.6 (2023-07-19)

Features Added

  • DALL-E image generation is now supported. See the Azure OpenAI quickstart for conceptual background and detailed setup instructions.
    • OpenAIClient gains a new GetImageGenerations method that accepts an ImageGenerationOptions and produces an ImageGenerations via its response. This response object encapsulates the temporary storage location of generated images for future retrieval.
    • In contrast to other capabilities, DALL-E image generation does not require explicit creation or specification of a deployment or model. Its surface as such does not include this concept.
  • Functions for chat completions are now supported: see OpenAI's blog post on the topic for much more detail.
    • A list of FunctionDefinition objects may be populated on ChatCompletionsOptions via its Functions property. These definitions include a name and description together with a serialized JSON Schema representation of its parameters; these parameters can be generated easily via BinaryData.FromObjectAsJson with dynamic objects -- see the README for example usage.
    • NOTE: Chat Functions requires a minimum of the -0613 model versions for gpt-4 and gpt-3.5-turbo/gpt-35-turbo. Please ensure you're using these later model versions, as Functions are not supported with older model revisions. For Azure OpenAI, you can update a deployment's model version or create a new model deployment with an updated version via the Azure AI Studio interface, also accessible through Azure Portal.
  • (Azure OpenAI specific) Completions and Chat Completions responses now include embedded content filter annotations for prompts and responses
  • A new Azure.AI.OpenAI.AzureOpenAIModelFactory is now present for mocking.

Breaking Changes

  • ChatMessage's one-parameter constructor has been replaced with a no-parameter constructor. Please replace any hybrid construction with one of these two options that either completely rely on property setting or completely rely on constructor parameters.

1.0.0-beta.5 (2023-03-22)

This is a significant release that brings GPT-4 model support (chat) and the ability to use non-Azure OpenAI (not just Azure OpenAI resources) to the .NET library. It also makes a number of clarifying adjustments to request properties for completions.

Features Added

  • GPT-4 models are now supported via new GetChatCompletions and GetChatCompletionsStreaming methods on OpenAIClient. These use the /chat/completions REST endpoint and represent the OpenAI Chat messages format.
    • The gpt-3.5-model can also be used with Chat completions; prior models like text-davinci-003 cannot be used with Chat completions and should still use the GetCompletions methods.
  • Support for using OpenAI's endpoint via valid API keys obtained from https://platform.openai.com has been added. OpenAIClient has new constructors that accept an OpenAI API key instead of an Azure endpoint URI and credential; once configured, Completions, Chat Completions, and Embeddings can be used with identical calling patterns.

Breaking Changes

A number of Completions request properties have been renamed and further documented for clarity.

  • CompletionsOptions (REST request payload):
    • CacheLevel and CompletionConfig are removed.
    • LogitBias (REST: logit_bias), previously a <string, int> Dictionary, is now an <int, int> Dictionary named TokenSelectionBiases.
    • LogProbability (REST: logprobs) is renamed to LogProbabilityCount.
    • Model is removed (in favor of the method-level parameter for deployment or model name)
    • Prompt is renamed to Prompts
    • SnippetCount (REST: n) is renamed to ChoicesPerPrompt.
    • Stop is renamed to StopSequences.
  • Method and property documentation are broadly updated, with renames from REST schema (like n becoming ChoicesPerPrompt) specifically noted in <remarks>.

1.0.0-beta.4 (2023-02-23)

Bugs fixed

  • Addressed issues that sometimes caused beta.3's new GetStreamingCompletions method to execute indefinitely

1.0.0-beta.3 (2023-02-17)

Features Added

  • Support for streaming Completions responses, a capability that parallels setting stream=true in the REST API, is now available. A new GetStreamingCompletions method on OpenAIClient provides a response value StreamingCompletions type. This, in turn, exposes a collection of StreamingChoice objects as an IAsyncEnumerable that will update as a streamed response progresses. StreamingChoice further exposes an IAsyncEnumerable of streaming text elements via a GetTextStreaming method. Used together, this facilitates providing faster, live-updating responses for Completions via the convenient await foreach pattern.
  • ASP.NET integration via Microsoft.Extensions.Azure's IAzureClientBuilder interfaces is available. OpenAIClient is now a supported client type for these extension methods.

Breaking Changes

  • CompletionsLogProbability.TokenLogProbability, available on Choice elements of a Completions response value's .Choices collection when a non-zero LogProbability value is provided via CompletionsOptions, is now an IReadOnlyList<float?> vs. its previous type of IReadOnlyList<float>. This nullability addition accomodates circumstances where some tokens produce expected null values in log probability arrays.

Bugs Fixed

  • Setting CompletionsOptions.Echo to true while also setting a non-zero CompletionsOptions.LogProbability no longer results in a deserialization error during response processing.

1.0.0-beta.2 (2023-02-08)

Bugs Fixed

  • Adjusted bad name finishReason to finish_reason in deserializer class

1.0.0-beta.1 (2023-02-06)

Features Added

  • This is the initial preview release for Azure OpenAI inference capabilities, including completions and embeddings.