AgentTravel: Knowledge-Augmented LLM Agent Framework for Urban Travel Planning

🚩 This is the repo of our paper summited to NeurIPS 2025 Workshop NORA.

🛠️ Requirements

• LLaMA Factory for finetuing LLMs

git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e ".[torch,metrics]"

• vLLM for startting local LLMs services

pip install vllm

📜 Quick Start

1. Download the database we processed and packaged in Google Drive, then Unzip to the root directory of this project.

2. Prepare CityGPT-T

   # finetune 
   cd /usr/exp/CityGPT-Travel/train/beijing
   bash run_qwen2.5_sft.sh
   
   # lora merge
   CUDA_VISIBLE_DEVICES=0,1 llamafactory-cli export ./examples/merge_lora/qwen2.5_lora_sft.yaml
   
   # start service
   cd /usr/exp/CityGPT-Travel
   CUDA_VISIBLE_DEVICES=0 vllm serve ./model_zoo/beijing/qwen2.5-7b/lora_merged --port 23199 --dtype auto --api-key sk-token-2025 --enable-auto-tool-choice --tool-call-parser hermes

3. Run TravelAgent

   python run_agent.py --city_en=beijing --platform=vLLM --model_name=citygpt-t-beijing

4. Conduct TravelBench

   python ./travel_bench/know_eval.py --city_en=beijing
   python ./travel_bench/trip_eval.py --city_en=beijing