haute 0.1.9

Haute — Open-source pricing engine for insurance teams on Databricks

Haute

Open-source pricing engine for insurance teams on Databricks.

Build, visualise, and deploy pricing pipelines as Python code — with a browser-based GUI that stays in sync.

pip install haute

---

What is Haute?

Haute gives insurance pricing teams a code-first, GUI-friendly way to build rating pipelines. Write standard Python with Polars, see it instantly in a visual editor, and deploy to a live API with one command.

• Build pipelines in code or the GUI — both stay in sync
• Run the same pipeline for 1-row live quotes and million-row batch jobs
• Deploy to Databricks MLflow Model Serving with haute deploy

Python code is always the source of truth. The GUI is a live, editable view.

---

Quick Start

1. Install

pip install haute

# For deployment to Databricks:
pip install haute[databricks]

2. Create a project

mkdir my_project && cd my_project
haute init

This scaffolds everything you need in the current directory:

haute.toml              ← project & deploy config
.env.example           ← Databricks credentials template
main.py                ← starter pipeline
data/                  ← your data files
test_quotes/           ← JSON payloads for pre-deploy testing

3. Write a pipeline

# main.py
import polars as pl
import haute

pipeline = haute.Pipeline("motor_pricing", description="Motor premium calculation")


@pipeline.node(path="data/policies.parquet", deploy_input=True)
def policies() -> pl.DataFrame:
    """Read policy data — this is the live API input."""
    return pl.scan_parquet("data/policies.parquet")


@pipeline.node(external="models/freq.cbm", file_type="catboost", model_class="regressor")
def frequency_model(policies: pl.DataFrame) -> pl.DataFrame:
    """Predict claim frequency."""
    df = policies.with_columns(
        freq_pred=pl.Series(obj.predict(policies.select("Area", "VehPower", "DrivAge").to_numpy()))
    )
    return df


@pipeline.node
def calculate_premium(frequency_model: pl.DataFrame) -> pl.DataFrame:
    """Calculate the technical premium."""
    return frequency_model.with_columns(
        premium=(pl.col("freq_pred") * 500).round(2)
    )


@pipeline.node(output=True)
def output(calculate_premium: pl.DataFrame) -> pl.DataFrame:
    """Final output returned by the API."""
    return calculate_premium

4. Run it

haute run

5. Open the GUI

haute serve

This opens a browser-based visual editor where you can:

• Drag and drop nodes from a palette
• Connect them with edges to define data flow
• Write Polars code in each transform node
• Click any node to preview its output data
• Toggle API Input on a data source to mark it as the live input
• Hit Run to execute the full pipeline
• Hit Save to write back to .py

6. Deploy

cp .env.example .env     # fill in your Databricks credentials
haute deploy

That's it. Your pipeline is now a live API on Databricks Model Serving.

---

Deployment

Haute deploys your pipeline as an MLflow model on Databricks Model Serving. One command, no DevOps.

How it works

1. Marks — you tag one data source as deploy_input=True (the live API input) and one node as output=True (the API response)
2. Prunes — Haute traces backwards from the output node and deploys only the scoring path. Training data branches, sinks, and exploratory nodes are automatically excluded
3. Bundles — model files (.cbm, .pkl, etc.) and static data are packaged as MLflow artifacts
4. Validates — every JSON file in test_quotes/ is scored through the pruned pipeline before deployment. If anything fails, deployment is blocked
5. Deploys — the pipeline is logged as an MLflow pyfunc model and registered in the Model Registry

Configuration

All deploy settings live in haute.toml (committed to git):

[project]
name = "motor-pricing"
pipeline = "main.py"

[deploy]
target = "databricks"
model_name = "motor-pricing"
endpoint_name = "motor-pricing"

[deploy.databricks]
experiment_name = "/Shared/haute/motor-pricing"
catalog = "main"
schema = "pricing"
serving_workload_size = "Small"
serving_scale_to_zero = true

[test_quotes]
dir = "test_quotes"

Secrets go in .env (gitignored):

DATABRICKS_HOST=https://adb-xxxxx.azuredatabricks.net
DATABRICKS_TOKEN=your_token_here

Test quotes

Put JSON files in test_quotes/ with example requests. These are scored before every deploy:

[
  {"IDpol": 99001, "VehPower": 7, "DrivAge": 42, "Area": "C", "VehBrand": "B12"}
]

Dry run

Validate everything without actually deploying:

haute deploy --dry-run

  ✓ Loaded config from haute.toml
  ✓ Parsed pipeline (12 nodes, 14 edges)
  ✓ Pruned to output ancestors (5 nodes)
  ✓ Collected 2 artifacts (freq.cbm, sev.cbm)
  ✓ Inferred input schema (10 columns)
  ✓ Test quotes: single_policy.json     1 rows  ok  (18ms)
  ✓ Test quotes: batch_policies.json    5 rows  ok  (24ms)
  ✓ Validation passed
  Dry run complete — no model was deployed.

Calling the deployed API

curl -X POST https://<workspace>.databricks.net/serving-endpoints/motor-pricing/invocations \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"dataframe_records": [{"Area": "A", "VehPower": 5, "DrivAge": 35}]}'

---

Key Concepts

Pipelines

A pipeline is a DAG of decorated Python functions. Each function is a node that takes DataFrames in and returns a DataFrame out.

pipeline = haute.Pipeline("my_pipeline")

@pipeline.node
def transform(read_data: pl.DataFrame) -> pl.DataFrame:
    return read_data.filter(pl.col("age") > 25)

Edges

Edges define data flow. Function parameter names match upstream node names:

pipeline.connect("read_data", "transform")
pipeline.connect("transform", "output")

Fan-out / Fan-in

One node can feed multiple downstream nodes, and a node can receive multiple inputs:

@pipeline.node
def joined(claims: pl.DataFrame, exposure: pl.DataFrame) -> pl.DataFrame:
    return claims.join(exposure, on="IDpol", how="left")

pipeline.connect("claims", "joined")
pipeline.connect("exposure", "joined")

Scoring

The same pipeline code works for batch and live scoring:

# Batch: run the full pipeline
result = pipeline.run()

# Live: score a single row (same code path as the deployed API)
row = pl.DataFrame({"Area": ["A"], "DrivAge": [35]})
prediction = pipeline.score(row)

---

Node Types

Data Source

Reads data from a file. No code needed — just configure the path.

@pipeline.node(path="data/policies.parquet", deploy_input=True)
def policies() -> pl.DataFrame:
    return pl.scan_parquet("data/policies.parquet")

• deploy_input=True — marks this source as the live API input for deployment
• Supported formats: Parquet, CSV, JSON

Transform

The workhorse node. Write Polars code to filter, join, aggregate, or reshape data.

@pipeline.node
def frequency_set(policies: pl.DataFrame, claims: pl.DataFrame) -> pl.DataFrame:
    return policies.join(claims, on="IDpol", how="left")

In the GUI, two shorthand syntaxes are available:

• Chain syntax — start with . to chain off the first input: .filter(pl.col("Area") == "A").select("IDpol", "premium")
• Expression syntax — reference multiple inputs by name: policies.join(claims, on="IDpol", how="left")

External File

Load a model or config file, then use it in your code. The loaded object is available as obj.

@pipeline.node(external="models/freq.cbm", file_type="catboost", model_class="regressor")
def frequency_model(policies: pl.DataFrame) -> pl.DataFrame:
    df = policies.with_columns(
        freq_pred=pl.Series(obj.predict(policies.select("Area", "VehAge").to_numpy()))
    )
    return df

Type	Extension	How it loads
Pickle	.pkl	pickle.load()
JSON	.json	json.load()
Joblib	.joblib	joblib.load()
CatBoost	.cbm	CatBoostClassifier / CatBoostRegressor

Output

Marks the final node whose result becomes the API response:

@pipeline.node(output=True)
def output(calculate_premium: pl.DataFrame) -> pl.DataFrame:
    return calculate_premium

Data Sink

Writes data to disk. Sinks are pass-through during normal runs — writing only happens when you click Write in the GUI.

@pipeline.node(sink="output/frequency.parquet", format="parquet")
def frequency_write(frequency_set: pl.DataFrame) -> pl.DataFrame:
    return frequency_set

---

GUI

The visual editor runs in your browser at http://localhost:5173.

Area	Description
Left palette	Drag node types onto the canvas
Center canvas	Visual DAG with drag, zoom, connect
Right panel	Configure the selected node
Bottom panel	Data preview — click any node to see its output

Nodes marked deploy_input=True show a green API badge. Toggle it on/off in the node's config panel.

Code ↔ GUI sync

Everything round-trips:

• Edit in the GUI → saves back to .py
• Edit the .py in your text editor → GUI picks it up on next load
• Custom imports, helper functions, and constants are preserved in both directions

---

Pipeline Imports & Helpers

Every pipeline starts with import polars as pl and import haute. Add extra imports or helper functions via the Imports button (⚙) in the GUI toolbar, or write them directly in the .py file between the standard imports and the first @pipeline.node.

import numpy as np
from catboost import CatBoostClassifier

DISCOUNT_RATE = 0.95

def apply_discount(df, col):
    return df.with_columns(pl.col(col) * DISCOUNT_RATE)

---

CLI Reference

Command	Description
haute init	Scaffold a new project in the current directory
haute run [file]	Execute a pipeline and print results
haute serve	Start the visual editor
haute deploy [file]	Deploy the pipeline as a live API
haute deploy --dry-run	Validate and score test quotes without deploying
haute status [model]	Check the status of a deployed model

haute serve options

Flag	Default	Description
--host	127.0.0.1	Host to bind to
--port	8000	Backend API port
--no-browser	off	Don't auto-open the browser

haute deploy options

Flag	Description
--model-name	Override model name from haute.toml
--dry-run	Validate and score test quotes without deploying

---

Project Structure

After haute init, your project looks like:

haute.toml                ← project & deploy config (committed)
.env.example             ← Databricks credentials template (committed)
.env                     ← actual credentials (gitignored)
.gitignore
main.py                  ← pipeline code (source of truth)
main.haute.json           ← GUI layout state (node positions)
data/                    ← data files (.parquet, .csv)
test_quotes/             ← JSON payloads for pre-deploy validation
  example.json

• .py files are the source of truth — diffable, reviewable, testable
• .haute.json files store GUI layout (node positions) — not execution logic
• haute.toml is the single config file for project settings and deployment

---

Design Principles

1. Code is the source of truth — the .py file is the pipeline. The GUI is a view.
2. Same pipeline, every context — the same code runs for 1-row live quotes and million-row batch jobs.
3. Real Python, real Polars — no proprietary formula language. Your skills transfer.
4. Git-native — pipelines are plain files. Diff, review, branch, merge.
5. One-command deploy — haute deploy handles pruning, bundling, validation, and MLflow registration.
6. Testable — every node is a plain function. pytest just works.

---

Requirements

• Python >= 3.11
• For deployment: pip install haute[databricks] (adds MLflow + Databricks SDK)
• Works on Linux, macOS, and Windows

---

License

MIT