How rllm-team/rllm works
rLLM is a PyTorch library for relational table learning, combining standardized graph, table, and LLM-oriented modules to build and train learning methods. A typical user workflow is to select a dataset, represent and transform its relational data, then use loaders and model components for learning. Major branches also include standalone graph/table architectures and LLM-assisted feature engineering and retrieval. The README documents these capabilities; sampled source confirms the storage, LLM adapter, retrieval, and feature-engineering details, but not the wiring of every model or data stage.
Main entry points: retrieval_llm.py Retrieval prediction · feat_engineer.py Feature engineering · langchain_llm.py LangChain adapter · feat_llm.py FeatLLM training
Components 21
Data and loading
- Datasets
rllm/datasets - Graph and tables
rllm/data - Data loaders
rllm/dataloader - Neighbor sampling
rllm/dataloader/sampler
Preparation
- Preprocessing
rllm/preprocessing - Table transforms
rllm/transforms/table_transforms - Graph transforms
rllm/transforms/graph_transforms
Learning models
- Feature encoders
rllm/nn/encoder - Table networks
rllm/nn/conv/table_conv - Graph networks
rllm/nn/conv/graph_conv - RTL and task models
rllm/nn/models - Learning losses
rllm/nn/loss
LLM workflows
- Feature engineering
rllm/llm/llm_module/featllm/feat_engineer.py - FeatLLM training
rllm/llm/llm_module/featllm/feat_llm.py - Retrieval prediction
rllm/llm/llm_module/retrieval/retrieval_llm.py - Table retriever
rllm/llm/llm_module/retrieval/retriever.py - LLM interface
rllm/llm/llm_module/general_llm.py - LangChain adapter
rllm/llm/llm_module/langchain_llm.py
Shared runtime
- Data storage
rllm/data/storage.py
Other components
- Library user
- Language model
Connections 8
- Library user → Datasets: selects data
- Library user → RTL and task models: chooses method
- FeatLLM training → Feature engineering: generates features
- Retrieval prediction → Table retriever: retrieves examples
- Retrieval prediction → Language model: queries
- Feature engineering → Language model: queries
- LLM interface → LangChain adapter: adapts through
- LangChain adapter → Language model: invokes