v0.20.0
版本发布时间: 2024-06-25 23:18:11
llm-workflow-engine/llm-workflow-engine最新发布版本:v0.22.0(2024-09-26 03:42:49)
This release introduces support for data caching, including data caching for plugins.
Some provider plugins now cache their available models after retrieving the data via an API call.
Plugins can now be reloaded using the /plugin reload [plugin_name] command. For provider plugins that cache available models, this will update the available models to the latest ones available via the API.
If backend_options.title_generation.provider is being used for a custom title provider, the new backend_options.title_generation.model configuration option will allow selecting the provider model as well.
:fire_engine:Breaking Changes:fire_engine:
The plugin caching/reloading functionality required a refactor of some plugin architecture -- make sure to update to the latest code for all plugins at the same time as the update to this release.
Commit log
- Tue Jun 25 2024: formatting cleanup
- Tue Jun 25 2024: add doc for reloading plugins
- Tue Jun 25 2024: add help for reloading plugins
- Tue Jun 25 2024: clear cache when reloading plugin
- Mon Jun 24 2024: fallback on non-standard response object
- Mon Jun 24 2024: standardize provider data caching, overriding provider settings in config
- Mon Jun 24 2024: add cache_delete method to CacheManager
- Sun Jun 23 2024: add /plugin command, plugin reload API
- Sun Jun 23 2024: fix docstring
- Sun Jun 23 2024: fix user help doc
- Sun Jun 23 2024: simple cache manager, expose to plugins
- Sat Jun 22 2024: add backend_options.title_generation.model
- Sat Jun 22 2024: log full LLM response object
- Sat Jun 22 2024: update FakeMessagesListChatModel to match latest in PR
- Fri Jun 07 2024: more robust cleanup of tool definitions, by dereferencing
- Fri Jun 07 2024: add debug log for non-streaming LLM attributes
- Thu Jun 06 2024: formatting cleanup
- Thu Jun 06 2024: use full template env in build_message_from_template()
- Thu Jun 06 2024: gpt-4o -> gpt-4-turbo for coding preset As evidenced by https://scale.com/leaderboard/coding and personal use, gpt-4o is worse at coding than gpt-4-turbo at this time.
- Mon Jun 03 2024: fix provider set_value() to support defaults explicitly set to None