Ollama v0 35 und 0 35 1 Update Image © OllamaOllama v0 35 und 0 35 1 Update (Image © Ollama)

Transition from Generative to Decision-Based Outputs

The integration of decision models enables the automation of tasks that involve classification rather than content generation. This makes the platform suitable for industrial applications such as automated ticket triage, model routing, and the systematic classification of content. By returning probabilities instead of prose, these models provide a measurable confidence level for each decision.

Currently, two models are available for this specific functionality:

  • Nimble, developed by Bespoke Labs.
  • Tev1, developed by Together AI.

Implementing the SystemOne API

The operational workflow for decision models involves providing a specific state (the context) and one or more structured questions to the API. In a technical support scenario, for example, the state would be the content of a support ticket, and the question would prompt the model to classify the ticket into predefined categories such as billing, errors, or account access. The API’s response includes the selected option, the individual probabilities for each available option, and an overall confidence score. This structured format allows developers to set thresholds for automated actions; for example, a ticket can be automatically assigned only if the confidence score exceeds a certain percentage.

The API supports three different query types to meet various analysis needs:

  • 1. Selection: The model selects an option from a predefined list and returns the probability for each option.
  • 2. True/False: The model returns a probability value indicating whether a specific condition is true.
  • 3. Score: The model provides a numerical score based on an ordered set of criteria.

System Improvements and Bug Fixes

Version 0.35 includes several optimizations regarding user experience and system stability:

  • Optimization of the Settings Menu: The settings interface now opens immediately, without having to wait for the model recognition process to complete.
  • Fixes to the macOS User Interface: Fixes have been made to the macOS update menu and the application icon to ensure that available updates are displayed correctly at startup.
  • Stability during MLX download: A critical issue that caused the download of MLX models to hang indefinitely has been resolved.
  • Deprecated parameter: The typical_p parameter is now deprecated. To prevent system crashes, requests containing this parameter now generate a warning log instead of failing completely.