Gemini Decision Model

DecisionModel asks typed questions about a state and gets calibrated probabilities back, with no generated text to parse. This example uses Google’s gemini-decision-1 to decide what to do with a customer’s return request, then checks an attached photo.

Setup

pip install -U swarms
export GEMINI_API_KEY="..."

The model name picks the provider: any name that starts with gemini-decision goes to Gemini’s Decisions API. There is no google-genai dependency, since the client calls the HTTP API directly.

The example

import base64
from pathlib import Path

from swarms import DecisionModel

# Reads GEMINI_API_KEY and calls Gemini's /v1beta/decisions endpoint.
model = DecisionModel(model_name="gemini-decision-1")

request = {
    "message": (
        "The blue mug I ordered on Monday arrived cracked on Wednesday. "
        "I'd like a replacement, not a refund."
    ),
    "order": {"id": "8821", "age_days": 2, "total_usd": 24.0},
    "customer": {"tier": "Gold", "returns_last_90_days": 0},
}

result = model.run(
    state=request,
    questions={
        "resolution": {
            "type": "choice",
            "instructions": "Which resolution fits this return request?",
            "criteria": {
                "replace": "Send the same item again.",
                "refund": "Return the payment to the customer.",
                "reject": "The request is outside the return policy.",
                "other": "Anything outside these outcomes.",
            },
        },
        "severity": {
            "type": "score",
            "instructions": "How badly is the item affected?",
            "criteria": [
                "Cosmetic: the item works as intended.",
                "Impaired: the item works with a workaround.",
                "Unusable: the item cannot be used.",
            ],
        },
        "is_damaged": {
            "type": "noul",
            "instructions": "Did the item arrive damaged?",
            "criteria": {
                "true": "The item itself is broken or faulty.",
                "false": "A preference, sizing, shipping or pricing complaint.",
            },
        },
        "within_window": {
            "type": "noul",
            "instructions": "Is order.age_days 30 or less?",
        },
    },
)

answers = result["answers"]
resolution = answers["resolution"]

print(
    f"Resolution: {resolution['choice']} "
    f"(confidence {resolution['confidence']:.2f})"
)
print(f"Severity: {answers['severity']['score']:.2f} on a 0-2 scale")
print(f"Damaged: {answers['is_damaged']['noul']:.2f}")
print(f"Within window: {answers['within_window']['noul']:.2f}")

approved = (
    resolution["confidence"] >= 0.7
    and resolution["choice"] in ("replace", "refund")
    and answers["is_damaged"]["noul"] >= 0.8
    and answers["within_window"]["noul"] >= 0.8
)
print(
    f"Action: {resolution['choice']} automatically"
    if approved
    else "Action: send to a human"
)

# A list of content blocks carries text, images, audio, video and PDFs.
image_path = Path(__file__).parents[2] / "images" / "new_logo.png"
image_base64 = base64.b64encode(image_path.read_bytes()).decode(
    "ascii"
)

has_logo = model.noul(
    state=[
        {"type": "text", "text": "Photo attached to request #8821."},
        {
            "type": "image",
            "data": image_base64,
            "mime_type": "image/png",
        },
    ],
    instructions="Does the photo show a company logo?",
)
print(f"Photo shows a logo: {has_logo:.2f}")

print(f"Cost of both requests: {model.calculate_cost()}")

How it works

One request, four questions. run() sends the request and all four questions together. Each question is answered on its own against the same state, so adding a question does not change the other answers.
Structured state. The state is a dict, which is sent to Gemini as JSON text. Instructions can name its fields by key path, as within_window does with order.age_days.
Three question types. choice picks one option, score rates the state against ordered levels, and noul returns the probability that a statement is true.
Yes/no descriptions. The true and false criteria on is_damaged say what each side means. DecisionModel renames them to the keys Gemini expects.
Thresholds in code. The request is approved automatically only when the resolution is confident and both policy checks pass. Everything else goes to a person.
Media. A list of content blocks passes through to Gemini unchanged, so the second call sends a text block and a base64 image. The same shape works for audio, video and PDFs.
Cost. calculate_cost() prices the tokens of both requests. Gemini bills input tokens only.

Output

The values depend on the model. The shape is:

Resolution: replace (confidence 0.97)
Severity: 1.85 on a 0-2 scale
Damaged: 0.98
Within window: 0.99
Action: replace automatically
Photo shows a logo: 0.96
Cost of both requests: {'input_tokens': 612, 'output_tokens': 0, 'input_cost': 6.12e-05, 'output_cost': 0.0, 'total_cost': 6.12e-05}

Yes/no probabilities from Gemini fall between 0.01 and 0.99, so a threshold above 0.99 never matches.

Switching provider

The questions and answers have the same shape for every provider, so the text part of this example runs on another decision model by changing the name:

model = DecisionModel(model_name="gpt-6-luna")  # OpenAI, reads OPENAI_API_KEY
model = DecisionModel()  # TypeSafe jev-latest, reads TYPESAFE_API_KEY

The list of content blocks is Gemini’s format. OpenAI takes images as a list of user messages instead; see the reference.

DecisionModel reference

Every provider, method and question type

Source

The example file in the swarms repo

Index

Basic Examples

Model Providers

Single Agent

Tree of Thoughts

Multi-Agent Examples

Comms Channels

Auto Agent Builder

Scheduled Agents

Applications

Research

Technical Analysis

Use Cases

Finance

Voice Agents

Integrations

MCP Tool Integrations

Deployment

CLI

Gemini Decision Model

Setup

The example

How it works

Output

Switching provider

DecisionModel reference

Source

​Setup

​The example

​How it works

​Output

​Switching provider

​Related

DecisionModel reference

Source

Setup

The example

How it works

Output

Switching provider

Related