Overview
Decisio is an open-source serving layer for structured decisions. It accepts text plus typed questions in the System One wire format, then returns a probability for every option, using one forward pass per question and generating no text.
The service runs on user-controlled hardware and uses frozen official model weights without project-specific training. It is independent from TypeSafe and is designed for applications that need machine-readable choices rather than free-form responses.
Key Features
- Handles prompts up to 32,768 tokens and as many as 255 options per question.
- Supports choice, yes-or-no, and scored decision structures.
- Serves Gemma 4 31B, Qwen3.6 35B A3B, or Gemma 4 12B through the same routes.
- Provides calibration, task registration, and prefix caching across supported bases.
- Runs on a GPU server, Mac, or laptop depending on the selected model build.
How It Works
Applications send a state containing the source text and define the questions to answer. Decisio evaluates each question and returns probabilities across its allowed outcomes, enabling software to select an answer or apply a confidence threshold.
Low-confidence results can be escalated to a person or a larger model. The project recommends choosing thresholds from the application’s own labelled examples because probability quality depends on the questions being asked.
Integrations and Use Cases
Decisio targets structured automation such as support triage, support routing, browser agents, and real-time control demos. Existing System One integrations include packages for LangChain, Vercel AI SDK, n8n, TanStack AI, and Pipecat, although those integrations are maintained by third parties.
Its local serving model is suited to workflows that require predictable typed output, probability-aware routing, and control over inference hardware rather than a hosted per-token API.