CFT-RAG-2025

CFT-RAG: An Entity Tree Based Retrieval Augmented Generation Algorithm With Cuckoo Filter

1. Code Usage

Configuration

pip install uv
uv sync
pip install lab-1806-vec-db==0.2.3 python-dotenv sentence-transformers openai pybloom_live
export HF_ENDPOINT=https://hf-mirror.com
python -m spacy download zh_core_web_sm

Complete Process of CFT-RAG

Arguments:

• vec-db-key: The key for the vector database.
• tree-num-max: The maximum number of trees to build.
• entities-file-name: The name of the entities file.
• search-method: The search method to use:
  • 0 for Vector Database Only
  • 1 for Naive Tree-RAG
  • 2 for Bloom Filter Search in Tree-RAG
  • 5 for improved Bloom Filter Search in Tree-RAG
  • 7 for Cuckoo Filter in Tree-RAG
  • 8 for Approximate Nearest Neighbors in Tree-RAG
  • 9 for Approximate Nearest Neighbors in Graph-RAG
• node-num-max: The maximum number of nodes to build.

Example:

python main.py --tree-num-max 50 --search-method 7

Test Cuckoo Filter

Individually testing the performance of the improved Cuckoofilter and the sorting results:

python test_tree.py

2. Reference

TRAG-cuckoofilter is based on https://github.com/efficient/cuckoofilter.

Use of datasets:

Dataset	MedQA	AESLC	DART	Rui'an People's Hospital
Scale	Large	Medium	Medium	Small
Source	https://github.com/jind11/MedQA	https://huggingface.co/datasets/Yale-LILY/aeslc	https://github.com/Yale-LILY/dart	https://www.rahos.gov.cn