LiR3AG

• AAAI 2026 A Lightweight Rerank Reasoning Strategy Framework for Retrieval-Augmented Generation

Environmental Installation

conda create -n lir3ag python=3.10
conda activate lir3ag
pip install -r requirements.txt

Quick Start

Datasets

In this paper, we use the following datasets for evaluation:

• HotpotQA
• 2WikiMultiHopQA
• MultiHopRAG
• MuSiQue

Prerequisites

• A prepared document index with FAISS (if your workflow uses vector retrieval).
• Input dataset in JSON or JSONL format (fields typically: id, question / query; corpus items: id, text / content / passage).

Basic usage examples

Run the reproduce example that calls Ollama (example):

python LiR3AG/reproduce/rag/qwen3.py \
  --input "../../query/HotpotQA.jsonl" \
  --output "../data" \
  --index_dir "../../index/HotpotQA" \
  --model qwen3:14b \
  --top_k 5 \
  --think

Output

• Outputs are saved as JSON / JSONL files.
• Typical output fields: id, question, think (intermediate chain-of-thought or reasoning), response (final answer), and optionally retrieved_docs or ranked_passages.

Tips

• Ensure the index has been generated before running retrieval experiments.
• Start with a small subset of data to tune hyperparameters (e.g., --top_k, prompt templates, model selection).
• Use descriptive output directories (experiment name + timestamp) to keep results organized.

For more examples and parameter descriptions, inspect the scripts under LiR3AG/scripts and LiR3AG/reproduce.

Citation

@article{chen2025lir,
  title={LIR $\^{} 3$ AG: A Lightweight Rerank Reasoning Strategy Framework for Retrieval-Augmented Generation},
  author={Chen, Guo and Huang, Junjie and Xie, Huaijin and Sun, Fei and Jia, Tao},
  journal={arXiv preprint arXiv:2512.18329},
  year={2025}
}