adi-decision

ADI — Agent Decision Intelligence

Rank options, quantify confidence, and explain the decision.

ADI is a Python library for structured multi-criteria decisions. Give it options, weighted criteria, constraints, and evidence confidence; it returns a ranked result with explainability, sensitivity analysis, and feedback-driven learning.

It is designed for two use cases:

• application code that needs auditable decision logic
• LLM agents that should call a deterministic ranking tool instead of improvising tradeoffs

Status: alpha. The core API is usable today, but expect refinement as real-world usage expands.

Why ADI

• Deterministic ranking: scores and ranks options from explicit structured inputs.
• Confidence-aware output: low-confidence evidence can reduce score impact or lower final confidence.
• Explainable decisions: returns criterion contributions, constraint reports, and counterfactual guidance.
• Scenario analysis: supports what-if runs and cross-scenario comparison.
• Learning loop: feedback can update saved preference profiles for future runs.
• Agent-ready: ships OpenAI and Anthropic tool schemas out of the box.

Install

From source:

git clone https://github.com/dimgouso/adi-Agent-Decision-Intelligence.git
cd adi-decision
pip install -e .

With development tooling:

pip install -e ".[dev]"

After PyPI publishing is enabled:

pip install adi-decision

Quick Start

from adi import DecisionRequest, decide
from adi.schemas.decision_request import Criterion, Option, OptionValue

request = DecisionRequest(
    options=[
        Option(
            name="Supplier A",
            values=[
                OptionValue(criterion_name="cost", value=100),
                OptionValue(criterion_name="quality", value=0.80, confidence=0.95),
            ],
        ),
        Option(
            name="Supplier B",
            values=[
                OptionValue(criterion_name="cost", value=85),
                OptionValue(criterion_name="quality", value=0.68, confidence=0.70),
            ],
        ),
    ],
    criteria=[
        Criterion(name="cost", weight=0.45, direction="cost"),
        Criterion(name="quality", weight=0.55, direction="benefit"),
    ],
    policy_name="balanced",
)

output = decide(request)

print(output.best_option)
for item in output.ranking:
    print(item.option_name, item.score, item.confidence)

LLM Agent Integration

from adi import call_adi_tool, get_openai_tools

tools = get_openai_tools()

# Pass `tools` to your OpenAI client.
# When the model calls `adi_decide`, route the arguments here:
result = call_adi_tool(tool_arguments)

CLI And API

CLI:

adi validate request.json
adi decide --input request.json
adi decide --input request.json --policy risk_averse

API:

uvicorn adi.interfaces.api:app --reload

Key endpoints:

• GET /health
• POST /decide
• GET /policies
• GET /policies/{policy_name}

Core Features

• Policies: balanced, risk_averse, and exploratory
• Constraints: hard elimination or soft penalties
• Sensitivity: robustness checks under weight perturbation
• Confidence handling: explicit per-value confidence or evidence-derived confidence
• Feedback: accept, reject, or override actions update user profiles
• Typed schemas: Pydantic models for request and response contracts

Development

Run the full local verification suite:

pip install -e ".[dev]"
ruff check adi/
pytest
python -m build
twine check dist/*

Release And Publishing

• Changelog: CHANGELOG.md
• Publishing guide: docs/PUBLISHING.md
• GitHub Releases: github.com/dimgouso/adi-Agent-Decision-Intelligence/releases

License

MIT