Public API

Import the package as import relflow as rf. This map points to the contracts behind the common workflow; constructor options live with their owning type.

Model And Schema

API	Purpose
rf.Model(...)	Build a Lightning module from named fields and branches. An existing rf.Schema is also accepted.
rf.Branch(...)	Group child fields into a repeated local context.
rf.Attention, rf.Mean	Choose the branch reduction; reduction=None preserves routed tokens.
rf.Mask, rf.Jitter	Configure visibility/objectives and continuous input perturbation.
rf.Address, rf.where, rf.predicate	Address and select schema nodes.
model.select, update, extend, delete, reset, override	Inspect or change the schema.

See Model Tree, Data Types, and Schema Mutation.

Data And Training

API	Purpose
rf.ArrowDataModule	Named splits backed by Arrow tables, files, datasets, or restartable Arrow factories.
rf.PolarsDataModule	Eager DataFrames converted once to the Arrow pipeline.
rf.source	Configure Arrow file reading.
rf.preprocess, rf.Preprocessor	Transform eager Polars frames before binding data.
rf.adamw	Configure the AdamW optimizer factory, fused by default.
rf.RollbackCheckpoint	Save validation checkpoints and restore the best one at fit end.
model.compile	Attach compilation to selected encoders and attention pools; returns the same model.

CustomDataModule and SyntheticDataModule support application-owned mapping streams. Prefer the Arrow and Polars modules when the data already has that form. See Data Modules, Preprocessing, Training, Pretraining And Fine-Tuning, and Performance.

Prediction And Persistence

API	Purpose
model.save, rf.Model.load	Save and restore the schema and learned model state.
model.predict	Return predictions as an Arrow table.
model.encode, model.forward, model.write	Explicit encoding, Torch computation, and Arrow writing stages.
rf.Writer	Write Lightning prediction batches to rank-partitioned Parquet files.
rf.postprocess, rf.Postprocessor	Transform eager Polars frames after prediction writing.
rf.Deployment	Serve models through the optional HTTP runtime.

Serving types, including Accelerator, JSONBackend, and ModelSource, require relflow[serving]. See Prediction Output, Postprocessors, and Serving.

Extensions

rf.Extension registers the request, tensorfield, embedder, decoder, output, and related behavior for a datatype. RequestBase, TensorFieldBase, EmbedderBase, DecoderBase, TensorInput, Context, and RaggedField provide its shared contracts. See Custom Tensorfields.