NVIDIA OO Agents Decisions
Turn typed agent methods into calibrated Jev routing and triage decisions without a chat model.
Added to Jevfast
How it uses Jev
In the decision-model quickstart, a support message is the state for typed questions declared in Python return annotations. DecideStrategy compiles a bool or Enum (or a Pydantic model containing them) into one decision request, sends it through the configured DecisionClient, and rebuilds the declared return value. The triage example asks Jev to select a support department and decide whether the issue needs immediate action; application code reads those typed values and prints the triage result.
What Jev decides
Support-message triage (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · department
YOUR APP
INSTRUCTION
Select the team that owns the request.
STATE
The triage method receives message: "I was charged twice for my March invoice." DecideStrategy constructs state from that method parameter and the instruction to use only the supplied message.
JEV · CHOICE
- BILLING
- TECHNICAL
- ACCOUNT
Question 2 · urgent
YOUR APP
INSTRUCTION
Decide whether the request needs immediate action.
STATE
The triage method receives message: "I was charged twice for my March invoice." DecideStrategy constructs state from that method parameter and the instruction to use only the supplied message.
JEV · NOUL
YesNo
App workflow
- Configure the decision client
Read the model, endpoint, and API key from DECISION_MODEL, DECISION_ENDPOINT, and OPENROUTER_API_KEY, with quickstart defaults for the model and endpoint. Pass the client to SupportRouter as its decision_model.
- Declare typed decisions
Define the department Enum and its Criteria, then declare department, confident_department, and triage methods with DecideStrategy. Triage's Pydantic model declares both department Choice and urgent bool questions with instructions.
- Send each message
For every sample message, call the three generated methods. Each call sends its own decision request; the triage method sends both fields in one request.
- Use typed answers and close the client
Print the department and triage values. For the detailed department result, compare the selected option's probability with the 0.8 threshold and route low-probability cases to manual review. Close the DecisionClient in a finally block.
Why it is interesting
The framework turns ordinary Python types and field metadata into named, typed decision questions. One method can batch related questions into a single request, while an Enum and Criteria make the available choices and their meanings explicit in code. A separate detailed ChoiceDecision can retain probability evidence and apply a threshold before the application accepts a value.