This is the official repository for our paper "Stop-RAG: Value-Based Retrieval Control for Iterative RAG".

We recommend using uv with Python 3.11 for dependency management. To install the necessary dependencies, run:

pip install uv
uv sync --no-dev

To run Stop-RAG training and evaluate the results, follow the steps below. All experiments were run on 4 H100 GPUs.

There are three types of variables to set:

• DATASET: the name of the dataset to use (musique or hotpotqa or 2wikimultihopqa)
• RETRIEVER: the type of retriever to use (contriever or bm25)
• METHOD: the base pipeline to use (ours or corag)

First, download all datasets and build retrieval embeddings from the corpora.

./scripts/download.sh {RETRIEVER}

Datasets will be stored in data/raw/{DATASET}, and the retrieval corpora, embeddings and indexes will be saved in data/corpus/{DATASET}.

Run the chosen pipeline to prepare the training and evaluation data.

./scripts/dataset.sh {DATASET} {RETRIEVER} {METHOD}

The processed data will be saved in data/processed/{DATASET}/{METHOD}/{RETRIEVER}.

Now run the training script. WandB logging is enabled by default, so the environment variables WANDB_API_KEY and WANDB_ENTITY must be set. To disable WandB logging, set WANDB_DISABLED=true.

./scripts/train.sh {DATASET} {RETRIEVER} {METHOD}

To run evaluation, first compute the scores for all trained checkpoints and find the best checkpoint and threshold.

./scripts/stop_rag_find_best.sh {DATASET} {RETRIEVER} {METHOD}

This script will print the best checkpoint and threshold. Substitute these values into the following command to run the final evaluation.

./scripts/stop_rag_test.sh {DATASET} {RETRIEVER} {METHOD} {CKPT} {THRESHOLD}

You can also evaluate the LLM-Stop baseline for comparison.

./scripts/llm_stop_test.sh {DATASET} {RETRIEVER} {METHOD}

The code for Contriever is adapted from EfficientRAG, and the dataset download scripts are adapted from IRCoT.