Retrieval-augmented generation (RAG) greatly benefits language models (LMs) by providing additional context for tasks such as document-based question answering (DBQA). Despite its potential, the power of RAG is highly dependent on its configuration, raising the question: What is the optimal RAG configuration? To answer this, we introduce the RAGGED framework to analyze and optimize RAG systems. On the representative DBQA tasks, we study two classic sparse and dense retrievers, and four top-performing LMs in encoder-decoder and decoder-only architectures. Through RAGGED, we uncover that different models suit substantially varied RAG setups. While encoder-decoder models monotonically improve with more documents, we find decoder-only models can only effectively use <5 documents, despite often having a longer context window. RAGGED offers further insights into LMs' context utilization habits, where we find encoder-decoder models rely more on contexts and are thus more sensitive to retrieval quality, while decoder-only models tend to rely on knowledge memorized during training.
To recreate the conda environment, run
conda create -n ragged -y python=3.10
pip install -r requirements.txt
To run and evaluate the retriever, see retriver/README.md.
To run and evaluate the reader, see reader/README.md.
To conduct downstream RAGGED analysis, see analysis_framework/README.md.
Our datasets are available on Huggingface
Specify corpus_dir and corpus_name and see download_data.py for how to download and save the files in appropriate folders.
For Pubmed corpus for BioASQ, the corpus name is pubmed.
For KILT wikipedia corpus, the corpus name is kilt_wikipedia.
After downloading the datasets, process the corpus for ColBERT format by running python retriever/data_processing/create_corpus_tsv.py --corpus $corpus --corpus_dir $corpus_dir, which outputs $corpus_dir/${corpus}/${corpus}.json.
Specify data_dir and dataset_name and see download_data.py for how to download the file to ${data_dir}/${dataset_name}.jsonl.
We support Natural Questions (KILT ver), HotpotQA (KILT ver), and BioASQ11B.
The above files are ready for BM25, but not for ColBERT. To reformat them for ColBERT, run python retriever/data_processing/create_query_tsv.py --data_dir $data_dir --dataset $dataset, which outputs $data_dir/${dataset}-queries.tsv.
To adapt for BM25, format your corpus and query as jsonl files as instructed here. To adapt for ColBERT, format your corpus and query datasets as instructed here.
If you use our code, datasets, or concepts from our paper in your research, we would appreciate citing it in your work. Here is an example BibTeX entry for citing our paper:
@article{hsia2024ragged,
title={RAGGED: Towards Informed Design of Retrieval Augmented Generation Systems},
author={Jennifer Hsia and Afreen Shaikh and Zhiruo Wang and Graham Neubig},
journal={arXiv preprint arXiv:2403.09040},
year={2024}
}For any questions, feedback, or discussions regarding this project, please feel free to open an issue on the repository or contact us:
- Contact
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent