Version 1.69

Release Date: 17.09.2023

1. ENHANCEMENTS

• OpenAI Integration in Text-Clustering Training API: To improve the text-clustering model, a new "llm" parameter has been introduced to the training API, in addition to the existing configuration settings. This parameter is optional, ensuring it does not impact existing applications.

  • If the "llm" parameter included in the request body is given "openai", OpenAI's LLM (Large Language Model) is used during the training and the inference.

• Waking Up Sleeping Pods for Inference Requests: In this release, we've introduced a feature to automatically wake up inference pods that have been put to sleep due to timeout. These pods are recreated when a new inference request for text classification (MLP) is received. Importantly, if the train data remains the same, the inference pod is created without the need for retraining. This enhancement optimizes the system's performance and ensures the efficient use of resources.

• Activation of Authorization for AI Modules: To access and use AI modules through public model endpoints and swagger pages, users now must be authorized. This authorization requirement does not affect current integrations through APIs. This enhancement strengthens the security of your AI modules by preventing unauthorized requests from being sent directly to the endpoints.

• Deletion of Published Models for Inactive Tenants: Existing published models for inactive tenants are now automatically deleted. This helps to manage resources and maintain system efficiency by removing modules that are no longer in use due to tenant inactivity.

2. IMPROVEMENTS

• Restriction on modelId Parameter: The use of underscores in the "modelId" parameter sent in training requests is restricted. The "modelId" parameter can now only contain lowercase letters (a-z), numbers (0-9), and hyphens ("-"). Hyphens are also not allowed at the beginning and end of the text. This change ensures consistency and compatibility in the modelId values across the system.

• Assigning Labels to Pod Names: To enhance the clarity and manageability of the system, we've implemented the assignment of labels to pod names. Labels for training and published model pods are now assigned based on the product integration, modelId, and full tenant name. This update is especially valuable for understanding which product uses which ML pods.

• Bug Fix: An issue where an empty list was returned during file copying from an experiment to the published model folder in MinIO after training has been resolved. This ensures that the correct "inferenceStatus" after after training is completed.