Skip to content

Latest commit

 

History

History
132 lines (89 loc) · 4.49 KB

README.md

File metadata and controls

132 lines (89 loc) · 4.49 KB
Raw

A Self-Adaptive Planning Agent For Multimodal RAG

Pytorcharxiv badge

Repo for Benchmarking Multimodal Retrieval Augmented Generation with Dynamic VQA Dataset and Self-adaptive Planning Agent

🌏 The Chinese Web Demo is avaiable at ModelScope now！

  • We propose OmniSearch, a self-adaptive retrieval agent that plans each retrieval action in real-time according to question solution stage and current retrieval content. As far as we known, OmniSearch is the first planning agent for multimodal RAG.
  • We reveal that existing VQA-based mRAG benchmarks fail to reflect the feature that real-world questions require dynamic knowledge retrieval, and propose novel Dyn-VQA dataset, which contains three types of dynamic questions.
  • We benchmark various mRAG methods with leading MLLMs on Dyn-VQA, demonstrating their flaw in providing sufficient and relevant knowledge for dynamic questions.

💡 Perfomance

The performance of various MLLMs with different mRAG strategies are shown below:

More analysis experiments can be found in the paper.

📚 Dyn-VQA Dataset

The json item of Dyn-VQA dataset is organized in the following format:

{
    "image_url": "https://www.pcarmarket.com/static/media/uploads/galleries/photos/uploads/galleries/22387-pasewark-1986-porsche-944/.thumbnails/IMG_7102.JPG.jpg/IMG_7102.JPG-tiny-2048x0-0.5x0.jpg",
    "question": "What is the model of car from this brand?",
    "question_id": 'qid',
    "answer": ["保时捷 944", "Porsche 944."]
}

🔥 The Dyn-VQA will be updated regularly. Laset version: 202412.

🛠 Dependencies

pip install -r requirement.txt

Details

  • Python = 3.11.9
  • PyTorch (>= 2.0.0)
  • pillow = 10.4.0
  • requests = 2.32.3
  • google-search-results = 2.4.2
  • serpapi = 0.1.5

💻 Running OmniSearch

We have release the code of GPT-4V-based OmniSearch for English questions.

Before running, please replace with your own OPENAI key and Google_search key. OPENAI key is at 11-th line of main.py

GPT_API_KEY = "your_actual_key_here"
headers = {
    "Authorization": f"Bearer {GPT_API_KEY}"
}

Google_search key is at 10-th line of search_api.py

API_KEY = "your api-key"

The result is saved to the path:

output_path = os.path.join(meta_save_path, dataset_name, "output_from_gpt4v.jsonl")

Run the main.py file:

python main.py --test_dataset 'path/to/dataset.jsonl' --dataset_name NAME --meta_save_path 'path/to/results'

🔍 Evaluation

The evaluation script for token F1-Recall of the output answers can be used as follows:

python evaluate.py --evaluate_file_path [path to output jsonl file] --lang [language of the
 QA dateset: en/zh]

🔥 TODO

  • Release code for Qwen-VL-Chat based OmniSearch
  • Release the corresponding model weight
  • Create a benchmark for Dyn-VQA

📄 Acknowledge

  • The repo is contributed by Xinyu Wang, Shuo Guo, Zhen Zhang and Yangning Li.
  • This work was inspired by ReACT, SelfAsk, FleshLLMs. Sincere thanks for their efforts.

📝 Citation

@article{li2024benchmarkingmultimodalretrievalaugmented,
      title={Benchmarking Multimodal Retrieval Augmented Generation with Dynamic VQA Dataset and Self-adaptive Planning Agent}, 
      author={Yangning Li and Yinghui Li and Xinyu Wang and Yong Jiang and Zhen Zhang and Xinran Zheng and Hui Wang and Hai-Tao Zheng and Pengjun Xie and Philip S. Yu and Fei Huang and Jingren Zhou},
      year={2024},
      eprint={2411.02937},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2411.02937}, 
}

When citing our work, please kindly consider citing the original papers. The relevant citation information is listed here.