Devotd Github
NewAI agents

osolmaz/switch v0.1.1 — Choice-based decision model harness

osolmaz/switch is a decision model harness framework that connects finite choices across multi-stage workflows.

Repositoryosolmaz/switch ↗

osolmaz/switch is a decision model harness framework that connects finite choices across multi-stage workflows.

GitHub description: A framework for building decision-model harnesses for any kind of task. Your program offers the model a short list of next steps, and the model picks one at a time until it can answer.

stars
★ 343
Forks
23
Language
TypeScript
License
MIT
Latest release
v0.1.1
Read original source ↗View on GitHub ↗

osolmaz/switch is a TypeScript framework structured so that a model picks one of several candidates for the next step. The latest release published on the GitHub repository is v0.1.1. When the program presents a short list of candidates, the model selects them one at a time until it arrives at an answer. This repository is not about one specific task but about a decision model harness for multiple kinds of tasks.

The core is how a model that answers with finite choices is wired into a multi-stage workflow. Developers implement the states, the question generation, and the transition functions according to their domain. A state represents the current position within the task, question generation decides what to ask the model at that stage, and the transition function carries the model's choice into the next state. The README's document evaluation example uses this to locate the needed parts step by step and turn them into a verdict. It also introduces the ability to record execution traces and replay them without model calls. It is written in TypeScript and uses the MIT license. This design focuses on structuring tasks so that the model answers with predefined choices instead of freely producing long answers.

That said, the README states that the current model's evaluation capability is not enough to replace a strong conversational model. The scope of this framework therefore lies less in the model's evaluation ability itself than in the way tasks are structured through choices and state transitions.