Dense-Passage-Retrieval

Dense Passage Retrieval (DPR) is a highly regarded technique that underpins retrieval-augmented Large Language Models (LLMs). While the original DPR repository is an excellent resource for academic research, it offers a multitude of configurable options that may be challenging for beginners to navigate.

This repository presents a streamlined implementation of the DPR model using the Natural Questions dataset. It is designed to facilitate a clear and straightforward understanding of DPR without compromising on essential details. Additionally, a pre-trained DPR model is provided for immediate experimentation.

1. Requirements

NOTE: This repository is compatible with Python 3.10.

Install the appropriate PyTorch and FAISS version based on your CUDA version. Visit the PyTorch Previous Versions page and FAISS installation guide for detailed instructions.

conda install pytorch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 pytorch-cuda=12.4 -c pytorch -c nvidia
conda install -c pytorch -c nvidia faiss-gpu=1.9.0

Install the remaining Python libraries:

pip install transformers numpy pandas seaborn matplotlib tqdm wget

2. Installation

Clone this repository and navigate to its directory:

git clone https://github.com/YuITC/Dense-Passage-Retrieval.git
cd Dense-Passage-Retrieval

3. Download Data

This repository utilizes the following datasets:

• English Wikipedia Dump (Dec. 20, 2018): Serves as the source documents for answering questions.
• Natural Questions (NQ) Dataset: Provides the question-answer pairs for training and evaluation.

Execute the following commands to download the necessary data:

python utils/download_data.py --resource data.wikipedia_split.psgs_w100 --output_dir dpr-dataset
python utils/download_data.py --resource data.retriever.nq --output_dir dpr-dataset
python utils/download_data.py --resource data.retriever.qas.nq --output_dir dpr-dataset

4. Usage

This repository provides pre-trained query encoder and document encoder models, available here.

Embedding phase: Generate embeddings for queries and documents.

python embedder_main.py

Retrieving phase: Retrieve relevant documents based on queries.

# For retrieving on the dataset
python retriever_main.py --query_source dataset

# For retrieving on the demo queries
python retriever_main.py --query_source demo --demo_file demo/demo_actual_queries.json

5. Results

Retrieval Performance: The following table compares the retrieval performance of the replicated DPR model against the reported results in the original paper on the Natural Questions (NQ) dataset:

	Top-20	Top-100
Paper	78.4	85.4
Ours	79.1	86

Computational Costs: All experiments were conducted on an 2xT4 16GB GPU setup (powered by Kaggle).

	generate embedding	build & search index
Duration	15h	7m 27s

6. References

Dense Passage Retrieval for Open-Domain Question Answering by VVladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih.

facebookresearch/DPR – Official GitHub repository for DPR by Facebook Research.

Hannibal046/nanoDPR – A lightweight implementation of DPR.

YouTube Tutorial by @IRwithPUGGY – An instructional video explaining DPR concepts.

7. License

This project is licensed under the Apache License.

8. Contact

For any questions or feedback, please open an issue or contact lehuuphuoc2502yuitc@gmail.com.