Skip to content

Latest commit

 

History

History
207 lines (113 loc) · 7.85 KB

CHANGELOG.md

File metadata and controls

207 lines (113 loc) · 7.85 KB

Changelogs

v0.0.53 (2024-05-31)

  • Added support for new gpt-4-o model

  • Added support for external providers

  • Update certain validation behavior of Functions

  • TODO: Add batch create support

v0.0.52 (2024-02-28)

  • Added support for the following parameters in model_configurations in OpenAIManager:

    • ping_timeout - allows for custom timeouts for each client.

    • included_models - allows for more flexible setting of models in Azure.

    • weight - allows for weighted selection of clients.

  • Improved Healthcheck behavior to cache if successful for a period of time, and always recheck if not.

  • Added dimension parameter for embedding models.

v0.0.51rc (2024-02-07)

  • Modification of async_openai.types.context.ModelContextHandler to a proxied object singleton.

  • Begin adding support for external providers, such as together to allow usage in conjunction with OpenAI models. WIP.

  • Rework of api_resource and root_name in Route objects to be settable during initialization. This is to allow for flexibility for external providers.

  • Added capability to have multi-api-key support for external providers, allowing for automatic rotation between api keys.

v0.0.50 (2024-02-01)

Breaking Changes

  • The OpenAI client has been refactored to be a singleton ProxyObject vs a Type object.

    Currently, this API is accessible with async_openai.OpenAIManager, which provides all the existing functionality of the OpenAI client, with a few additional features.

    • OpenAIManager supports automatic proxy rotation and client selection based on available models.

    • OpenAIManager supports automatic retrying of failed requests, as well as enabling automatic healthchecking prior to each request to ensure the endpoint is available with auto_healthcheck_enabled, otherwise it will rotate to another endpoint. This is useful for ensuring high availability and reliability of the API.

    Future versions will deprecate the OpenAI client in favor of the OpenAIManager object.

  • Added new OpenAIFunctions class which provides a robust interface for creating and running functions. This class is also a singleton ProxyObject.

    This can be accessed through the OpenAIManager.functions object

v0.0.41 (2023-11-06)

Update to Latest OpenAI API

This version updates the API to the latest version of OpenAI's API, which includes the following changes:

  • addition of gpt-4-turbo models

  • Add additional supported parameters to chat endpoint. We maintain v1 parameters for azure endpoints, but will pass through the new parameters for openai endpoints.

  • Add gradual support for tools

Updates

  • Rework of validating models, which now is no longer done, and expects the user to pass the correct model name.

  • No longer supporting validate_max_tokens as there are now many different schemas for max_tokens depending on the model.

v0.0.40 (2023-10-18)

Potentially Breaking Changes

This version introduces full compatability with pydantic v1/v2 where previous versions would only work with pydantic v1. Auto-detection and handling of deprecated methods of pydantic models are handled by lazyops, and require lazyops >= 0.2.60.

With pydantic v2 support, there should be a slight performance increase in parsing pydantic objects, although the majority of the time is spent waiting for the API to respond.

Additionally, support is added for handling the response like a dict object, so you can access the response like response['choices'] rather than response.choices.

v0.0.36 (2023-10-11)

Additions

  • Added auto-parsing of pydantic objects in function_call parameters and return the same object schema in chat_response.function_result_objects.

v0.0.35 (2023-10-06)

Additions

  • Added auto_retry option to OpenAI client, which will automatically retry failed requests.
  • Added RotatingClients class which handles the rotation of multiple clients. This can be enabled by passing rotating_clients=True to the OpenAI client while configuring.
  • Added OpenAI.chat_create and OpenAI.async_chat_create methods which automatically handles rotating clients and retrying failed requests.
  • Added azure_model_mapping which allows automatically mapping of Azure models to OpenAI models when passing openai models as a parameter, it will automatically convert it to the Azure model. This is only done in chat implementation.

Fixes

  • Fixed api_version Configuration handling.
  • Fixed parsing of function_call in streaming implementation.

v0.0.34 (2023-10-06)

Changes

  • Updated default api_version to 2023-07-01-preview
  • Added __getitem__ attributes to completion and chat objects, allowing them to act like dict objects.
  • Added functions and function_call to Chat completion routes.
    • function.properties can pass through a pydantic object which will convert it automatically to a dict json schema.
  • Added function_call attribute in ChatMessage objects, allowing for easy access to the function call.
  • Streaming is not supported for functions at this time.

v0.0.33 (2023-08-24)

Changes

  • Updated auto-configuring httpx's logger to be disabled if debug_enabled is set to False.

v0.0.32 (2023-08-23)

Changes

  • Updated headers behavior and parameter, allowing it to be passed to each API call.

  • Updated auth behavior, which now utilizes httpx.Auth rather than injecting into the header directly.

  • Added custom_headers configuration that can be passed to the OpenAI client during initialization.

  • Added customization of connection_pool, controlling the number of concurrent connections to the API.

  • Reworked streaming implementations, which previously didn't properly work.

  • Added parse_stream parameter (default: true) which defers parsing of the stream util it is called with result.stream or result.astream, rather than parsing the stream as it is received.

v0.0.31 (2023-08-11)

Changes

  • Updated some behaviors of the OpenAI Client.

    • allow for customization of retry behavior or completely disabling it.
  • Routes now take the is_azure parameter during init rather than using @property to determine the route.

  • The _send method is better optimized for retry behaviors.

Fixes

  • Resolved model endpoint.
  • Resolved handling of Azure models.

v0.0.30 (2023-08-10)

Potentially breaking changes in this version

Changes

  • Refactored the architecture of the OpenAI Client to accomodate multi-client initialization. i.e. OpenAI can now be initialized with multiple API keys and will automatically rotate between them, as well as switch back and forth between Azure and OpenAI.

  • Settings are initialized after first call, rather than globally.

  • Routes, Clients are configured after first call, rather than during initialization.

Fixes

  • Resolved embedding endpoints.

Updates

  • Changed default api-version to 2023-03-15-preview

v0.0.22 (2023-06-14)

  • Update pricing to reflect OpenAI's new pricing model
    • gpt-3.5-turbo
    • text-embedding-ada-002
  • Bugfix for consumption and usage validation in chat models
  • Added support for gpt-3.5-turbo-16k
  • Modified handling of gpt-3.5-turbo's consumption pricing to reflect prompt and completion usage
  • Modified default Embedding model to be ada

0.0.17 (2023-04-12)

  • Add better support for chatgpt models and gpt-4
  • Better validation max_tokens

0.0.11 (2023-03-07)

  • Added support for GPT-3.5 Turbo through async_openai.OpenAI.chat
  • Refactored async_openai.OpenAI to utilize a metaclass rather than initalizing directly

0.0.7 (2023-02-02)

  • Refactor async_openai.types.options.OpenAIModel to handle more robust parsing of model names.

0.0.3 (2022-12-21)

  • Fix proper charge for babbage and ada models.