Pydantic AI Adds Native DecisionModel Support for Jev, OpenAI, and Google Backends
Pydantic AI introduced a unified DecisionModel abstraction mapping Python output types directly to Jev, System One servers, OpenAI Decisions, and Gemini Decision 1.
Python’s Pydantic AI framework has added first-class support for decision models through a new DecisionModel abstraction that translates standard Pydantic output schemas directly into typed decision-model requests.
Python types compile straight into decision primitives
Instead of hand-crafting JSON question payloads, developers define ordinary typed Python models and pass them to an agent. Pydantic AI automatically maps a bool field to a yes/no probability, a Literal or Enum to a single choice, a list of labels to independent yes/no checks, a bounded float to raw probability, and rubric definitions to an ordered score. Multi-field models execute in a single request, and tool or output-type selection becomes a typed routing step.
Four backends and automatic LLM fallback
Out of the box, Pydantic AI ships four decision backends behind the same interface: TypeSafeModel (typesafe:jev-latest), generic SystemOneModel for any /v1/systemone-compatible server such as Laya, CLM, or Ollama-hosted models, OpenAIDecisionsModel, and GoogleDecisionsModel. Swapping providers requires changing only the model string.
Crucially, Pydantic AI pairs DecisionModel with ModelFallback: whenever a decision model selects a route that requires free-form text generation—or returns low confidence on an ambiguous input—the framework automatically escalates that step to a full language model. That hybrid pattern lets cheap, fast decision calls handle routine branches while keeping an LLM on standby for open-ended work.