RAG with Passage Combination

This repository provides tools and scripts to run retrieval and train a model to retrieve passage combinations that helps downstream retrieval-augmented QA the most. A baseline is also provided by simply retrieving and taking the top k passages.

Usage

1. Run Retrieval:

   • Use the provided scripts indicated in retrieval README to run the retrieval process. This step involves generating representation vectors for the static documents dataset and retrieving the best matching passages given the query vectors.

2. Train Model:

   • Train a model to retrieve passage combinations using the provided .sh files.

   bash qa_passage_combination/train_and_evaluate.sh

3. Evaluate Model:

   • Evaluate the trained model using the provided .sh files.

   bash qa_passage_combination/test.sh

4. Baseline:

   • A baseline is provided by simply retrieving and taking the top k passages. This can be used as a reference to compare the performance of your trained model.

   bash qa_baseline/rerank.sh
   bash qa_baseline/run_qa_all.sh

Acknowledgement

This project is based on and inspired by the work and code from the following repositories:

1. Hannibal046/nanoDPR
2. castorini/pyserini
3. AI21Labs/in-context-ralm
4. StonyBrookNLP/ircot

We thank the authors of these repositories for their contributions to the community.

License

This project is licensed under the MIT License.