azure-ai-inference

Features Added

• Added a client for Image Embeddings, named ImageEmbeddingsClient. See package README.md and new samples.
• Added support for Chat Completions response message in JSON format that adheres to a given JSON schema. Also known as "structured output". See new samples sample_chat_completions_with_structured_output.py and sample_chat_completions_with_structured_output_pydantic.py.
• Made input argument content a positional argument (in addition to keyword argument), in the constructors of UserMessage, SystemMessage, AssistantMessage and ToolMessage. For example, you no longer need to write UserMessage(content="my message"). Simply write UserMessage("my message"). All samples were updated accordingly.

Breaking Changes

• If you previously configured your ChatCompletionClient.complete() call to output JSON format without a scheme, you have this in your code: response_format=ChatCompletionsResponseFormatJSON(). To maintain the same functionality, replace this with response_format="json_object". We however recommend that you now switch to output JSON format with a provided schema if your AI model supports it: response_format=JsonSchemaFormat(...).

Bugs Fixed

• Fix a bug that would cause an error when tracing was enabled and azure-core-tracing-opentelemetry was not installed and asynchronous chat completion was used.
• Enforce distinct timestamps on prompt and completion tracing events to preserve the order for chat history.

Features Added

• OpenTelemetry tracing:
  • Method AIInferenceInstrumentor().instrument() updated with an input argument enable_content_recording.
  • Calling AIInferenceInstrumentor().instrument() twice no longer results in an exception.
  • Added method AIInferenceInstrumentor().is_content_recording_enabled()
• Support Prompty and prompt template from string. PromptTemplate class outputs an array of dictionary with OpenAI compatible message format.

Bugs Fixed

• Fix tracing for asynchronous streaming.

Features Added

• Support for OpenTelemetry tracing. Please find more information in the package README.md.
• When constructing clients using input credential of type AzureKeyCredential, two HTTP request headers are sent simultaneously for authentication: Authentication: Beater <key> and api-key: <key> (previously only the first one was sent). This is to support different inference services, removing the need for the application to explicitly specify an additional HTTP request header.

Features Added

• Support chat completion streaming response with function arguments (tool calls). Add new classes StreamingChatResponseMessageUpdate and StreamingChatResponseToolCallUpdate.
• Support text embeddings result in base64 encoded string format.
• Nicely formated print of chat completions and embeddings result objects.

Breaking Changes

• Classes ChatCompletionsToolSelectionPreset, ChatCompletionsNamedToolSelection and ChatCompletionsFunctionToolSelection renamed to ChatCompletionsToolChoicePreset ChatCompletionsNamedToolChoice and ChatCompletionsNamedToolChoiceFunction respectively.
• Update the object type of embeddings property on EmbeddingsResult, from embedding: List[float] to embedding: Union[str, List[float]].
• Instead of base class ChatCompletionsToolCall and derived class ChatCompletionsFunctionToolCall, we now have a flat representation of only one class ChatCompletionsToolCall that that represents a function tool. This is because the only support tool is a function call.

Bugs Fixed

• Fix setting of chat completions response format, to allow response in JSON format. See classes ChatCompletionsResponseFormat (base class) and derived classes ChatCompletionsResponseFormatJSON and ChatCompletionsResponseFormatText.

Features Added

• Allow setting default chat completions configuration in the ChatCompletionsClient constructor.
• Allow setting default embeddings configuration in the EmbeddingsClient constructor.
• Add model as an optional input argument to the embed method of EmbeddingsClient.

Features Added

Add model as an optional input argument to the complete method of ChatCompletionsClient.

Breaking Changes

The field input_tokens was removed from class EmbeddingsUsage, as this was never defined in the REST API and the service never returned this value.

• Initial beta version