How cmavro/gnn-rag works
Project Type & Purpose
• A research‐oriented KGQA system combining:
• GNN pipeline (“gnn/”) to retrieve answer nodes & reasoning paths over a knowledge graph
• RAG pipeline (“llm/”) to verbalize paths and query LLMs for final answers
File Structure Overview
• gnn/: data loading, model definitions (GraftNet, NSM, ReaRev), training/evaluation scripts
• llm/: prompt templates, results storage, scripts, and a modular Python package:
– align_kg, joint_training, qa_prediction, llms (language model abstractions), utils
Key Architectural Patterns
• Pipeline pattern: sequential stages from data → GNN inference → prompt construction → LLM call → evaluation
• Modular design: clear separation of GNN vs. LLM concerns; pluggable encoders & GNN variants; abstract LLM interface
Main Components
• Data Loader (KG + QA dataset)
• GNN Engine
• Submodules: question encoder, KG reasoner, model zoo
• Retrieval Output (answer nodes + paths)
• Verbalizer / Prompt Builder
• LLM Client
• pluggable backends (ChatGPT, LLaMA, Flan‐T5, etc.)
• Evaluator & Results Store
Relationships & Data Flow
• Optional branches: training vs. inference; joint_training/align_kg sub-pipelines
Technologies & Frameworks
• Python, PyTorch (GNNs), HuggingFace transformers, FastChat API
• Shell scripts for orchestration