Choose a pipeline node by what you want to do

A pipeline is a recipe: start with observations, prepare useful features, evaluate a model on held-out observations, and save the resulting predictor. Each node performs one step. Click a node below for its worked recipe, explanatory figure, expected output and common mistakes.

New to pipelines? Read Operator steps: describe an algorithm and its settings, Preprocessing: change what the model sees, split: define what must be unseen and model: turn features into predictions first. Then add views, paths or searches as your question requires.

Start with the ordinary learning workflow

Your goal	Node	What changes
Describe an algorithm and its settings	Operator steps: describe an algorithm and its settings	Declares the operator
Normalize, smooth or reduce features	Preprocessing: change what the model sees	Feature values or columns
Define cross-validation observations	split: define what must be unseen	Which rows train/validate each fit
Predict concentration or class	model: turn features into predictions	Learns a predictor
Scale/encode targets and invert predictions	y_processing: learn on one scale, report on another	Target representation
Inspect spectra, targets and folds	Chart nodes: ask a concrete question of the current data	Adds a diagnostic figure

Add information without confusing rows and columns

Your goal	Node	What changes
Retain raw, normalized and derivative views	feature_augmentation: retain several views of the same observations	View axis
Join several transformed feature blocks	concat_transform: give one model several feature blocks	Column count
Create perturbed training observations	sample_augmentation: give training examples plausible variants	Training row count
Annotate unusual observations	tag: flag a sample and keep it	Metadata, no deletion
Remove flagged rows from training	exclude: choose which training rows influence the fit	Fitting population
Handle several scans of each material	Repetitions: several measurements are still one physical sample	Grouping or repetition representation

Compose more advanced experiments

Your goal	Node	What to learn next
Keep several preprocessing/model paths	branch: let the same samples follow different paths	Duplication, separation and source routing
Join features, model predictions or rows	merge: decide what to bring back together	Output axes and OOF stacking
Compare many alternative recipes	Generators: turn an experiment plan into concrete recipes	Two transforms × three complexities = six recipes
Learn errors left by a base model	residual: learn an additive correction	Base prediction plus correction
Adapt preprocessing to another instrument	auto_transfer_preproc: choose preprocessing for an instrument change	Adaptation cohort and recommendations

Complete spelling index

Spellings	Explanation
class, function, instance, direct object, import-path string	Operator steps: describe an algorithm and its settings
params, name, force_layout	Operator steps: describe an algorithm and its settings
model	model: turn features into predictions
split, direct splitter object	split: define what must be unseen
preprocessing, bare transformer, sequential subpipeline list	Preprocessing: change what the model sees
y_processing	y_processing: learn on one scale, report on another
feature_augmentation, action	feature_augmentation: retain several views of the same observations
concat_transform	concat_transform: give one model several feature blocks
sample_augmentation	sample_augmentation: give training examples plausible variants
tag, exclude	tag: flag a sample and keep it, exclude: choose which training rows influence the fit
branch, by_source, by_metadata, by_tag, by_filter	branch: let the same samples follow different paths
merge, merge_sources, merge_predictions	merge: decide what to bring back together
rep_to_sources, rep_to_pp, rep_fusion	Repetitions: several measurements are still one physical sample
residual	residual: learn an additive correction
auto_transfer_preproc	auto_transfer_preproc: choose preprocessing for an instrument change
chart_2d, chart_3d, y_chart, chart_y	Chart nodes: ask a concrete question of the current data
fold_chart, chart_fold, fold_*	Chart nodes: ask a concrete question of the current data
spectra_dist, spectral_distribution, spectra_envelope	Chart nodes: ask a concrete question of the current data
augment_chart, augmentation_chart, augment_details_chart, augmentation_details_chart	Chart nodes: ask a concrete question of the current data
exclusion_chart, chart_exclusion	Chart nodes: ask a concrete question of the current data
_or_, _range_, _log_range_, _grid_, _zip_, _cartesian_, _chain_, _sample_	Generators: turn an experiment plan into concrete recipes
pick, arrange, then_pick, then_arrange, count	Generator keywords: choices, combinations and their exact outputs
_seed_, _weights_, _mutex_, _requires_, _exclude_, _preset_	Generator keywords: choices, combinations and their exact outputs
_tags_, _metadata_, _depends_on_	Generator keywords: choices, combinations and their exact outputs (advanced/reserved limits)

Choose the execution environment once

A recipe file describes intent. The runtime must actually support its algorithms and topology. The worked tabs identify which route is used:

Route	Appropriate scope
Native language facade	Qualified finite method recipes in Python/R/Octave/WASM
Python dag-ml	Supported host operators and graph workflows
Python legacy	Historical host controllers explicitly selected in examples

These are not interchangeable claims of portability. Complex Python controller recipes cannot be made native simply by changing JSON keys. Start with 10. Run the same recipe in your language for complete examples in your language. Consult Native Capability Preflight or Public execution matrix when extending the topology.

Full schemas and internal parser details remain in Pipeline Keywords Reference. The Operator catalog: choose a job, then inspect the method lists algorithms, distinct from workflow nodes.