Overview

PyCaret 4.0 is three things stacked together:

1. The engine — a Python library that exposes a single OOP API for five ML task types. This is what you pip install.
2. The backend — an optional FastAPI service that fronts the engine over HTTP, persists runs to a database, and serves predictions.
3. The dashboard — an optional React UI that talks to the backend.

You can use the engine alone (the notebook workflow). You can run the backend without the dashboard (a stateless engine.run(config) endpoint). You can stand up all three for a full self-hosted MLOps cockpit.

The engine#

Every task type has an Experiment subclass:

• ClassificationExperiment
• RegressionExperiment
• ClusteringExperiment
• AnomalyExperiment
• TimeSeriesExperiment

They share a base class with a uniform verb surface — fit, create_model, compare_models, tune_model, predict_model, finalize_model, save_model, load_model. The verbs are written against sklearn.pipeline.Pipeline (or sktime.forecasting.compose. ForecastingPipeline for time-series), so every fitted output is a real, portable scikit-learn artefact.

The engine state lives in experiment._fit_state — a dict populated by fit() with the train/test splits, the fold generator, the model registry, and the metric registry. Every verb reads from this dict rather than reaching into a god-class. (The "god-class drain" from 3.x → 4.0 took eighteen sessions; you can read the autopsy in the release notes.)

The backend#

pycaret-server (in services/api/) is a FastAPI app with:

• Workspace / project / experiment / run CRUD.
• A run orchestrator (thread-pool) that subscribes to engine events.
• WebSocket fan-out for live run streams.
• A pipeline registry + deployment serving layer.
• LLM-backed advisory endpoints (failure debugger, deployment review, drift analyst, …).
• Plot rendering endpoints (/runs/{id}/plots/{kind}).

Deploy it on its own. Pin a specific engine version. Wire it to your existing auth (JWT-friendly).

The dashboard#

apps/web/ is a Vite + React app that consumes the backend. It ships eight workflow screens (Datasets / Setup / Run / Models / Trials / Predict / Pipelines / Deployments) plus seven dashboard screens introduced in 4.0 (Workspace home / Model card / Data profile / Forecast workbench / Drift / Prediction explorer / Model comparison).

The dashboard never imports the engine directly. Everything it shows comes from the backend over /api/v1/* and the WebSocket event stream.

What's optional vs. required#

If you only want to fit models in a notebook, you only need the engine. Everything else is opt-in.