Decider

Decider is a Python framework for building, serving, and inspecting decision pipelines as versioned, deployable micro-services. Define pipelines from plain Python functions, compose them with | and &, save them as versioned JSON configs, and serve them over HTTP — all with a single consistent API.

Table of Contents

• Introduction
• Installation
• Concepts
• Usage Examples
• CLI
• Contributing
• License

Introduction

Decider is built around a few core ideas:

• Functions as nodes — plain Python functions that accept and return polars.Expr are wired into executable DAGs automatically, with no decorators or registries required.
• Composable pipelines — modules chain with | (sequential) or merge with & (parallel union), making it easy to build complex pipelines from simple parts.
• Versioned configs — every pipeline is serialisable to JSON. Configs are versioned, loadable by the server at startup, and hot-swappable without redeployment.
• Pluggable extensions — new module types are registered into a discriminated union at runtime, so the server always knows how to reconstruct any module from its config.

Installation

# Core library and CLI
pip install decider

# With serving dependencies
pip install "decider[serve-starlette]"   # uvicorn + starlette
pip install "decider[serve-sanic]"       # sanic

# With the interactive graph visualiser
pip install "decider[visualise]"

# Everything
pip install "decider[all]"

With uv (recommended):

git clone https://github.com/capitecbankltd/dsp_north-polrs.git
cd dsp_north-polrs
uv sync --all-extras

Concepts

Modules

A module is the basic unit of computation. The simplest way to create one is generate_from_functions, which turns plain functions into an executable module:

• Function name → output column — def dti_ratio(...) produces a dti_ratio column.
• Parameter name → input column or sibling output — parameters are resolved from the input DataFrame or from the output of another function in the same module.
• config parameter → Pydantic model injection — declare config: MyConfig and the config fields are promoted onto the module itself.

Pipelines

Modules compose with two operators:

• | — sequential: step_a | step_b | step_c passes each step's output as the next step's input.
• & — union: module_a & module_b merges both modules into a single compilation pass, computing all columns in one frame scan.

Versioned Configs

Every module can be saved to a versioned JSON config and reconstructed from it:

await module.asave("main", config_manager)
await config_manager.save_version(overwrite=True)

The server reads the latest version on startup and watches for updates.

Extensions

Custom module types are registered with register_graph_module and auto-discovered by initialize_decider from an extension_path directory. This keeps the core library small while allowing domain-specific modules to be developed and shipped independently.

Usage Examples

Your first module

import polars as pl
from decider.modules.functional import generate_from_functions

def dti_ratio(debt: pl.Expr, income: pl.Expr) -> pl.Expr:
    return debt / income

def credit_score(dti_ratio: pl.Expr) -> pl.Expr:
    return pl.lit(800) - dti_ratio * 200

Scorer = generate_from_functions("credit_scorer", dti_ratio, credit_score)
scorer = Scorer(name="scorer")

df = pl.DataFrame({"debt": [25_000.0], "income": [50_000.0]})
result = scorer({"input": df})
# shape: (1, 4) — debt, income, dti_ratio, credit_score

Config injection

from pydantic import BaseModel

class ScorerConfig(BaseModel):
    dti_weight: float = 200.0
    score_base: float = 800.0

def credit_score(dti_ratio: pl.Expr, config: ScorerConfig) -> pl.Expr:
    return pl.lit(config.score_base) - dti_ratio * config.dti_weight

Scorer = generate_from_functions("credit_scorer", dti_ratio, credit_score)
scorer = Scorer(name="scorer", dti_weight=150.0)  # config fields on the module

Sequential pipeline

features = FeatureModule(name="features")
scorer   = ScorerModule(name="scorer")
flags    = FlagModule(name="flags")

pipeline = features | scorer | flags
result = pipeline({"input": df})

Join then score

from decider.modules.primitives.join import JoinModule

join = JoinModule(name="enrich", left="transactions", right="users", on="user_id", how="left")
pipeline = join | scorer

result = pipeline({"transactions": txns_df, "users": users_df})

Save and load

import asyncio
from decider.config.file import JsonFileConfigManager
from decider.modules import GraphModule

mgr = JsonFileConfigManager(basepath="./configs")
asyncio.run(scorer.asave("main", mgr))
asyncio.run(mgr.save_version(overwrite=True))

# Reconstruct from disk
fresh = JsonFileConfigManager(basepath="./configs")
loaded = asyncio.run(fresh.get_latest())
module = GraphModule.model_validate(loaded.config["main"]).root

CLI

# Scaffold a new module
decider template module CreditScorer

# Scaffold a module into a shareable package
decider template module CreditScorer --package mylib

# Scaffold a new project
decider template project fraud_detection

# Start a server
decider serve                            # starlette on :8080
decider serve --engine sanic --workers 4
decider serve --port 9000 --reload

# Launch the interactive graph visualiser
decider visualise --project-dir projects/loan_scoring

Jupyter magic

%load_ext decider.magics

%%module CreditScorer

class CreditScorerConfig(BaseModel):
    weight: float = 200.0

def dti_ratio(debt: pl.Expr, income: pl.Expr) -> pl.Expr:
    return debt / income

def credit_score(dti_ratio: pl.Expr, config: CreditScorerConfig) -> pl.Expr:
    return pl.lit(800) - dti_ratio * config.weight

This writes decider_extensions/credit_scorer/__init__.py, reloads it, and injects CreditScorer into the notebook namespace. Add --package mylib to write into a proper uv src-layout package instead.

Contributing

We welcome contributions to Decider! Please refer to our Contributing Guide for full details.

• Fork the repository and create a branch from main.
• Install dependencies with uv sync --all-extras.
• Run tests with uv run pytest.
• Submit a Pull Request with a clear description of your changes.

License

This project is licensed under the MIT License. See the LICENSE file for details.

---

Thank you for your interest in Decider! We look forward to your contributions and feedback.