What is the 'decision' modality?
The decision output modality is a classification used by Coolhand for models that produce a structured, choice-based response instead of generating freeform text, images, or audio. While most large language models (LLMs) generate content, decision models are designed to evaluate inputs against a defined schema and return a specific, structured answer.
This makes them useful for tasks like:
- Classification
- Scoring and rating
- Extracting specific information
Example: TypeSafe Jev
The primary example of a model with the decision output modality is TypeSafe's Jev model. When you send a request to Jev, you provide a state (the data to be evaluated) and questions (the schema and criteria for evaluation). The model's response is not a generated text string, but a structured JSON object containing the answers.
For example, you might provide a user comment as the state and ask the model to classify its sentiment as 'positive', 'negative', or 'neutral' in the questions. The model would then return a structured response indicating its choice.
Because these models do not generate text in the same way as conversational LLMs, Coolhand categorizes their output modality as decision to distinguish them from generative models.