This repository releases the source code for the paper:

• Corrective Retrieval Augmented Generation.
  Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, Zhen-Hua Ling
  

Large language models (LLMs) inevitably exhibit hallucinations since the accuracy of generated texts cannot be secured solely by the parametric knowledge they encapsulate. Although retrieval-augmented generation (RAG) is a practicable complement to LLMs, it relies heavily on the relevance of retrieved documents, raising concerns about how the model behaves if retrieval goes wrong. To this end, we propose the Corrective Retrieval Augmented Generation (CRAG) to improve the robustness of generation. Specifically, a lightweight retrieval evaluator is designed to assess the overall quality of retrieved documents for a query, returning a confidence degree based on which different knowledge retrieval actions can be triggered. Since retrieval from static and limited corpora can only return sub-optimal documents, large-scale web searches are utilized as an extension for augmenting the retrieval results. Besides, a decompose-then-recompose algorithm is designed for retrieved documents to selectively focus on key information and filter out irrelevant information in them. CRAG is plug-and-play and can be seamlessly coupled with various RAG-based approaches. Experiments on four datasets covering short- and long-form generation tasks show that CRAG can significantly improve the performance of RAG-based approaches.

• 2024-10-08: Revise the prompt format in the inference stage to improve the performance in RAG and CRAG modes and make the generator follow instructions better.
• 2024-03-04: Release the knowledge preparation including knowledge refinement and knowledge search to gather correct, incorrect and ambiguous knowledge.
• 2024-03-02: Release the fine-tuning of the evaluator and training data preprocessed on PopQA.
• 2024-03-01: Release the data preprocess and data preparation for the inference of Self-CRAG.
• 2024-02-22: Release the inference of CRAG and the weights of the retrieval evaluator used in our experiments. Will release the inference of Self-CRAG and the fine-tuning of the retrieval evaluator soon.

Note: We use Python 3.11 for CRAG To get started, install conda and run:

git clone https://github.com/HuskyInSalt/CRAG.git
conda create -n CRAG python=3.11
...
pip install -r requirements.txt

• Download the eval_data created by Self-RAG (Asai et al., 2023) on PopQA, PubQA, Bio and Arc_challenge with retrieved results
• Download the LLaMA-2 fine-tuned by Self-RAG (Asai et al., 2023).
• Download the fine-tuned weights of the retrieval evaluator used in our experiments.
• Download the training data for the evaluator.

Run the following command to preprocess the dataset for questions and retrieval results. Specifically for PopQA, the label of each (question, passage) pair is also collected.

bash run_data_preprocess.sh

Run the following command to fine-tune the evaluator.

bash run_evaluator_training.sh

The training data is shared and can be downloaded, the method of label collection is similar to the test set preparation in scripts/data_process.py.

Run the following command to gather knowledge for inference, including correct, incorrect and ambiguous.

bash run_knowledge_preparation.sh

Specifically, you can also run the following commands individually.

According to the paper, we decompose the retrieval results and filter out irrelevant parts. Three modes are listed to decompose passages: fixed_num, excerption and selection. fixed_num segments passages into a fixed number of words, 'excerption' segments passages based on the end of the sentences, while passages are not divided in selection mode. You can choose the mode by --decompose_mode.

python internal_knowledge_preparation.py \
--model_path YOUR_EVALUATOR_PATH \
--input_queries ../data/$dataset/sources \
--input_retrieval ../data/$dataset/retrieved_psgs \
--decompose_mode selection \
--output_file ../data/$dataset/ref/correct 

Question rewriting and web searching are proposed here, thus an openai_api_key and a search_key are required. In this experiment, we utilized a third-party Google Search API platform for searching. Two selective modes including wiki and all are available. wiki visits pages related to Wikipedia preferentially, while all visit all pages equally.

python external_knowledge_preparation.py \
--model_path YOUR_EVALUATOR_PATH \
--input_queries ../data/$dataset/sources \
--openai_key $OPENAI_KEY \
--search_key $SEARCH_KEY \
--task $dataset --mode wiki\
--output_file ../data/$dataset/ref/incorrect 

Run the following command to combine both correct and incorrect knowledge for ambiguous action.

python combined_knowledge_preparation.py \
--correct_path ../data/$dataset/ref/correct \
--incorrect_path ../data/$dataset/ref/incorrect \
--ambiguous_path ../data/$dataset/ref/ambiguous 

Run the following command for CRAG inference.

bash run_crag_inference.sh

Run the following command for Self-CRAG data preparation.

bash run_selfcrag_preparation.sh

With this command, the retrieval results of the original input files of Self-RAG will be replaced by correct, incorrect and ambiguous context. Then follow the instructions at Self-RAG (Asai et al., 2023) for the ultimate results.

For Bio evaluation, please follow the instructions at the FactScore (Min et al., 2023) official repository.

python -m factscore.factscorer --data_path YOUR_OUTPUT_FILE  --model_name retrieval+ChatGPT --cache_dir YOUR_CACHE_DIR --openai_key YOUR_OPEN_AI_KEY --verbose

It is worth mentioning that, previous FactScore adopted text-davinci-003 by default, which has been deprecated since 2024-01-04 and replaced by gpt-3.5-turbo-instruct. Both results of CRAG and Self-CRAG reported are based on the text-davinci-003, which may differ from the current gpt-3.5-turbo-instruct evaluation.

For the other datasets, run the following command.

bash run_eval.sh

e.g., PopQA

python eval.py \
  --input_file eval_data/popqa_longtail_w_gs.jsonl \
  --eval_file ../data/popqa/output/YOUR_OUTPUT_FILE \
  --metric match 

PubHealth

python eval.py \
  --input_file eval_data/health_claims_processed.jsonl \
  --eval_file ../data/pubqa/output/YOUR_OUTPUT_FILE \
  --metric match --task fever

Arc_Challenge

python run_test_eval.py \
  --input_file eval_data/arc_challenge_processed.jsonl \
  --eval_file ../data/arc_challenge/output/YOUR_OUTPUT_FILE \
  --metric match --task arc_c

If you think our work is helpful or use the code, please cite the following paper:

@article{yan2024corrective,
  title={Corrective Retrieval Augmented Generation},
  author={Yan, Shi-Qi and Gu, Jia-Chen and Zhu, Yun and Ling, Zhen-Hua},
  journal={arXiv preprint arXiv:2401.15884},
  year={2024}
}