OpenDecision

OpenDecision answers typed questions about text, JSON, documents, and images. It runs locally and returns structured values.

Measured on CPU with short inputs and the default models. Longer inputs and simultaneous requests can use more memory. Memory details.

Text context is 512 tokens for Choice, Score, and Relation, or 8,192 tokens for yes/no (Noul), including the question text sent to the model. Tokens are pieces of words. Images are resized to 224 × 224 pixels, with 64 tokens per caption. The document API splits longer documents into passages. Input limits.

Text decisions

Route a customer request

I was charged twice and I need this fixed before my meeting this afternoon.

Question	Answer
Which department should handle this?	billing
Is this request urgent?	Yes

if department == "billing" and urgent:
    action = "prioritize_billing_request"

API call: 231 ms for both questions.

Read an insurance claim

The sample claim includes policy facts, medical records, bills, and employer records.

Question	Answer
Are the medical expenses documented?	established
Is a serious injury documented?	refuted
Are documents still outstanding?	established

These insurance answers come from the saved claim benchmark.

More text and document examples · Document API

Image decisions

Ask questions about a photo and combine the answers into an action.

Wildlife alert

Question	Answer
Is an animal visible?	Yes
What kind of animal is visible?	bear

if animal_present and animal_type == "bear":
    action = "sound_alarm"

Returned action: sound_alarm.

API call: 318 ms for 2 questions.

Bin collection

Question	Answer
Is a bin visible?	Yes
What condition is the bin in?	overflowing
Is litter scattered on the floor?	Yes

if bin_present and bin_condition == "overflowing":
    action = "request_collection"

Returned action: request_collection.

API call: 234 ms for 3 questions.

Fence repair

Question	Answer
Is a fence visible?	Yes
What condition is the fence in?	damaged

if fence_present and fence_condition == "damaged":
    action = "create_repair_task"

Returned action: create_repair_task.

API call: 155 ms for 2 questions.

Times are medians of five local HTTP calls on an Apple M3 with the models loaded. Each payload received one warmup call. The timer includes sending the request and reading the response. Measurement details.

These are recorded results from SigLIP2. The actions are simulated. The rules use a yes/no threshold of 0.5 and require a winning Choice probability of at least 0.6.

See more images, questions, and answers · Image API and setup

Install

With pip:

pip install OpenDecision

With uv:

uv add OpenDecision

Continue to the get started guide

What it provides

Type	Use	Result
Choice	Select one option.	Option name and probabilities
Noul	Test one statement.	A score from 0 to 1
Score	Use an ordered scale.	Weighted score and probabilities
Relation	Compare evidence with a statement and its opposite.	supports, contradicts, unknown, or conflicted
Document decisions	Ask questions about text or JSON.	Answers and source passages
Image decisions	Ask questions about a photo.	Answers and rule decisions

See code examples for each primitive

Start here

• Get started: install the package, run a Python example, and start the API.
• Primitives: use Choice, Noul, Score, and Relation.
• Image examples: photos, questions, returned answers, and combined rules.
• Image decisions: install image support and send API requests.
• Document decisions: ask questions about long text or JSON and select a yes/no mode.
• Evidence and rules: rank evidence and combine facts with rules.
• Examples: run the Doom demo and review the insurance and GDPR examples.

Interfaces

Interface	Use
Python	Call OpenDecision in the same process as the application.
POST /v1/systemone	Send state and typed questions to the API.
POST /v1/documents/decide	Send a document and typed questions to the API.
TypeSafe-compatible endpoint	Use a compatible TypeSafe SDK client with a local server.

Basic example

from opendecision import OpenDecisionEngine

engine = OpenDecisionEngine()

result = engine.choice(
    state="The customer was charged twice.",
    instructions="Which team should handle this request?",
    criteria={
        "billing": "Payments, invoices, refunds, and duplicate charges",
        "technical": "Software bugs",
        "sales": "Pricing and purchases",
    },
)

print(result["choice"])
# billing

Doom demo

OpenDecision chooses actions for a bot in ViZDoom's Deadly Corridor. This is the Skill 5 recording.

Watch Skill 1 | Watch Skill 3 | Run the demo