ColdRAG

ColdRAG is the official implementation of our Retrieval-Augmented, Large Language Model (LLM) based cold-start recommendation system.
This repository contains the full Qwen-based ColdRAG pipeline implemented with vLLM for LLM inference and BAAI/bge-m3 for embedding-based retrieval.

Environment Setup

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

Starting vLLM Servers

ColdRAG requires two vLLM servers:

Qwen LLM Server

export VLLM_QWEN_URL=http://localhost:8000/v1
export VLLM_MAX_MODEL_LEN=131072

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-7B-Instruct \
  --port 8000 \
  --gpu-memory-utilization 0.8 \
  --tensor-parallel-size 1 \
  --max-model-len 131072 \
  --rope-scaling '{"type":"yarn","factor":4.0,"original_max_position_embeddings":32768}' \
  --download-dir "$CACHE_DIR"

Embedding Server

export VLLM_EMBED_URL=http://localhost:8001/v1

python -m vllm.entrypoints.openai.api_server \
  --model BAAI/bge-m3 \
  --port 8001 \
  --gpu-memory-utilization 0.1 \
  --tensor-parallel-size 1 \
  --download-dir "$CACHE_DIR"

Dataset & Knowledg Graph Setup

Download the processed dataset and RAG index from: https://drive.google.com/drive/folders/1QkwQugctMfLlBBhHixhjpbRf4AgsX91u?usp=sharing

Place them like this:

ColdRAG/
├── dataset/
│     └── Video_Games/...
└── rag_output_qwen/
      └── Video_Games/...

Notes

• If rag_output_qwen/Video_Games/ is empty, ColdRAG will first run knowledge graph construction (indexing) and then proceed to inference.
• If the RAG index exists, ColdRAG skips indexing and runs inference directly.

Running ColdRAG

python main.py --model ColdRAG_qwen --dataset Video_Games --core 15 --cand_size 100 --k 10 --batch_size 5

Output Files

After running ColdRAG, two output files are generated:

Prediction Log File

outputs/ColdRAG_Video_Games_core15.json

Evaluation File

outputs/ColdRAG_Video_Games_core15_eval.json