v1.2.0 - Update logic, SDK adapters & storing models locally

In this update we address one of the major pain points when dealing with such an array of different services: version management. Additionally we enriched our SDK, which now contains more adapters to a few of your favourite open-source ML frameworks.

New update logic

Version 1.2.0 brings you a new and improved update logic. Previously, there was a potential for version mismatches of the different services that make up refinery resulting in potentially unexpected behaviour. We fixed that now with a new version overview screen that lets you update all the relevant services together! If refinery detects an update, you will get a notification, which takes you to the version overview. You can also manually go into that overview by clicking on the version number in the bottom left corner of the UI.

Screenshot of the new version overview. Note: as this is a screenshot from development, the installed version numbers are higher than the ones that are available on GitHub. Normal users should see a screen where the "installed version" matches the "remote version".

Screenshot of the new version overview. Note: as this is a screenshot from development, the installed version numbers are higher than the ones that are available on GitHub. Normal users should see a screen where the "installed version" matches the "remote version".

New adapters in the refinery Python SDK

Our Python SDK also got new exciting features. The includes two new adapters, which allow you to access your projects in refinery directly from Python code. Besides the previous adapter to format the data for Rasa NLU, our SDK now also offers an adapter to use the framework Sklearn (includes XGBoost and LightGBM) as well as getting the data for fine-tuning with HuggingFace transformers.

Read our short blog article about this for a better idea of what that enables you to do.

Storing HuggingFace models locally (hosted version only)

Version 1.2.0 also offers a new premium feature for our hosted version. Models can now be saved to disk, which is especially useful if refinery is supposed to run on-premise and offline. This also speeds up the embedding creation process when uploading additional data to the project as we don't have to pull your model from HuggingFace first!

Minor changes

  • updated Qdrant to a new version for the provided security fixes