ollama-decision

Python SDK for Ollama's Decision API (/v1/systemone). Convert text into structured, typed responses without prose generation.

Installation

pip install ollama-decision

Quick start

from ollama_decision import OllamaSystem1, noul_question, choice_question, score_question

model = OllamaSystem1('nimble-v3:9b-q4_K_M', timeout=30)

Define questions

refund_q = noul_question('wants_refund', 'Does user demand a refund?') intent_q = choice_question('intent', 'Request category?', { 'billing': 'Payment issues', 'technical': 'Bugs and errors', 'other': 'Other' }) sentiment_q = score_question('sentiment', 'Emotional tone', [ 'Negative', 'Neutral', 'Positive' ])

Bind state and questions

model['ticket'] = "I want my money back, this is frustrating!" model << ('ticket', refund_q) << ('ticket', intent_q) << ('ticket', sentiment_q)

Execute

response = model('ticket') print(response.json())

Question types

Boolean question returning a probability (0.0 to 1.0).

noul_question(name, instruction)

q = noul_question('is_urgent', 'Is this request urgent?')
# Returns: {"noul": 0.95}

Single selection from a dictionary of options.

choice_question(name, instruction, criteria)

q = choice_question('department', 'Route to?', {
    'sales': 'New purchases',
    'support': 'Technical help',
    'billing': 'Payment issues'
})
# Returns: {"choice": "billing", "confidence": 0.87, "probabilities": {...}}

Ordered scale evaluation (must be sorted from lowest to highest).

score_question(name, instruction, criteria)

q = score_question('priority', 'Priority level', [
    'Low', 'Medium', 'High', 'Critical'
])
# Returns: {"score": 2.3, "confidence": 0.78, "probabilities": {...}}

OllamaSystem1 API

Methods & Operators

model = OllamaSystem1('nimble-v3:9b-q4_K_M') -- Constructor model[state_name] = text -- Bind input text to a state name. model << (state_name, question) -- Attach a question object to the specified state. model(state_name) -- Execute the API request and return the HTTP response object.

Supported Models

nimble-v3:9b-q4_K_M — Best for multilingual tasks and high accuracy. tev1:4b — Fast English classification. tev1:0.8b — Ultra-lightweight for constrained environments. ..etc

License

MIT