Ragnarök

Ragnarök is a battleground for the best retrieval-augmented generation (RAG) models!

📟 Instructions

Source Installation

Create a new conda environment and install the dependencies:

conda create -n ragnarok python=3.10 -y
conda activate ragnarok
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 # For CUDA 11.8
pip install -r requirements.txt

PyPI Installation

pip install pyragnarok

Let's generate!

For example, to run the command-r-plus model on the rag24.researchy-dev_large dataset with bm25 retrieval methods with augmented generation based on the top-20 segments, you can run the following command:

python src/ragnarok/scripts/run_ragnarok.py  --model_path=command-r-plus  --topk=20 \
  --dataset=rag24.researchy-dev  --retrieval_method=bm25 --prompt_mode=cohere  \
  --context_size=8192 --max_output_tokens=1024 

Or to run the gpt-4o model on the rag24.raggy-dev dataset with bm25 followed by rank_zephyr_rho retrieval methods, with augmented generation based on the top-5 segments, you can run the following command:

python src/ragnarok/scripts/run_ragnarok.py  --model_path=gpt-4o  --topk=100,5 \
    --dataset=rag24.raggy-dev  --retrieval_method=bm25,rank_zephyr_rho --prompt_mode=chatqa  \
    --context_size=8192 --max_output_tokens=1024  --use_azure_openai

Contributing

If you would like to contribute to the project, please refer to the contribution guidelines.

🦙🐧 Model Zoo

Most LLMs supported by VLLM/FastChat should additionally be supported by Ragnarök too, albeit we do not test all of them. If you would like to see a specific model added, please open an issue or a pull request. The following is a table of generation models which we regularly use with Ragnarök:

✨ References

If you use Ragnarök, please cite the following:

TODO

🙏 Acknowledgments

This research is supported in part by the Natural Sciences and Engineering Research Council (NSERC) of Canada.