Overview

HybridRAG-Bench is designed to separate retrieval quality from reasoning quality.

Pipeline

  1. Collect a time-framed corpus that is external to model pretraining.

  2. Build hybrid knowledge from documents and extracted KG relations.

  3. Generate reasoning-grounded QA pairs from explicit reasoning paths.

  4. Validate benchmark quality with automated checks.

Repository Structure

  • arxiv_fetcher/: Corpus acquisition and processing.

  • dataset/: Dataset construction utilities.

  • kg/: Knowledge graph storage, preprocessing, and updates.

  • question_gen/: Multi-type reasoning question generation.

  • inference/: Baselines and retrieval-augmented inference methods.

  • run/: End-to-end pipeline and evaluation entry points.

Quick Start

Run the main pipeline modules from the project root:

python -m run.run_kg_preprocess
python -m run.run_kg_embed
python -m run.run_kg_update
python -m run.run_qa --dataset [movie, sports] --model [io, rag, kg, our]