ViDRILL - Vietnamese Deep Retrieval in the Expansive Legal Landscape

🇻🇳 Vietnamese Document Retrieval System for VLSP 2025

👉 If you find this project interesting and useful, please give it a star ⭐ to support us!

📍 Overview

ViDRILL is an advanced Vietnamese document retrieval system developed for the VLSP 2025 competition. It combines traditional retrieval methods, including embedding-based models and BM25, to achieve high performance in information retrieval from Vietnamese corpora.

While our official VLSP 2025 pipeline does not yet incorporate LLM-based retrieval, ViDRILL supports fine-tuning LLMs for:

• Query rewriting
• Document selection (retrieval)
• Reranking

This design allows future integration of LLM-enhanced retrieval to improve accuracy and reasoning.

---

🎯 Features

• Multi-stage Retrieval Pipeline: Optimized multi-step search for maximum accuracy

• Hybrid Search Methods:

  • Dense retrieval using E5 and GTE embeddings
  • Sparse retrieval using BM25
  • Neural reranking with BGE rerankers

• LLM Fine-Tuning Support (for future pipeline integration):

  • Query rewriting for more effective search
  • Document selection from top-k retrieval results
  • PPO/GRPO-based reinforcement learning for self-guided search
  • Interleaved reasoning + search with ViSearch-R1

• Vietnamese Language Support: Optimized for Vietnamese text

• Scalable Architecture: Handles large corpora using Qdrant vector database

---

🏗️ Architecture

ViDRILL/
├── pipeline/           # Main retrieval pipeline (current VLSP 2025 pipeline does not use LLM)
├── training/           # Model training modules
│   ├── retrieval/      # Dense retrieval training
│   ├── rerank/         # Reranking model training  
│   ├── llm/            # LLM fine-tuning (SFT or GRPO)
│   ├── DeepRetrieval/  # LLM query rewriting training
│   └── ViSearch-R1/    # LLM reasoning + self-searching training
├── prepare_data/       # Data preprocessing utilities
├── eval/               # Evaluation scripts
└── config/             # Training and system configurations

---

🚀 Setup and Usage

1. Installation

# Clone the repository
git clone https://github.com/xndien2004/ViDrill.git
cd ViDRILL

# Install dependencies
pip install -r requirements.txt

2. Data Preparation

# Preprocess corpus and build database
cd prepare_data
bash scripts/main.sh

# Build vector database
python build_db_corpus.py

---

3. Training Models (Optional / Future LLM Pipeline)

Dense Retrieval Models

cd training/retrieval

# Choose one model to train:
bash e5-instruct.sh    # E5 Dense Retrieval
# bash gte.sh           # GTE Dense Retrieval
# bash bge-m3.sh        # BGE-M3 Dense Retrieval

Reranking Models

cd training/rerank

# Choose one reranking script:
# bash bge-reranker.sh
# bash bge-reranker-mini.sh

LLM Fine-Tuning (Optional)

Option 1: Supervised Fine-Tuning (SFT)

cd training/llm
bash scripts/train_sft.sh

Option 2: GRPO / PPO Reinforcement Learning

cd training/llm
bash scripts/train_grpo.sh

Notes:

• Supports using top-k retrieved documents to train LLMs for document selection.
• Current VLSP 2025 pipeline does not use these LLM models, but they are ready for future integration.

---

DeepRetrieval: Query Rewriting with PPO (Optional)

cd training/DeepRetrieval

# 1. Prepare Data
bash scripts/run_data_process.sh

# 2. Start Retrieval Server
bash scripts/run_retrieval_server.sh

# 3. Train LLM with PPO (GRPO)
bash scripts/train_ppo.sh

Notes:

• Designed for improving query rewriting and selection accuracy.
• Can be integrated into the pipeline in the future.

---

ViSearch-R1: Reasoning & Self-Searching LLMs (Optional)

cd training/ViSearch-R1

# 1. Prepare Data
bash scripts/run_data_process.sh

# 2. Start Retrieval Server
bash scripts/run_retrieval_server.sh

# 3. Train LLM with RL (PPO / GRPO / Reinforce)
bash scripts/train_ppo.sh

Notes:

• LLM learns to reason over retrieved documents and call search engines.
• Supports various RL algorithms and LLMs.
• Currently for experimental/future usage; not in the VLSP 2025 pipeline.

---

4. Running the Retrieval Pipeline (Current VLSP 2025)

Encode Corpus

cd pipeline/scripts
bash encode_corpus.sh

Run Retrieval

# Hybrid: E5 + Sentence-BERT + BM25
bash main_e5_sentence_bm25.sh

# Dense only: E5 + Sentence-BERT
bash main_e5_sentence.sh

---

5. Evaluation

cd eval
bash scripts/eval.sh

---

📊 Pipeline Components

Dense Retrieval

• E5 Multilingual
• GTE Models
• BGE-M3

Sparse Retrieval

• BM25
• Hybrid Scoring

Reranking

• BGE Reranker
• LLM Reranking / PPO Self-Searching (future support)

Query Processing

• HyDE
• Query Expansion & Rewriting (future LLM support)

---

🛠️ Configuration

• ds_zero1.json - DeepSpeed ZeRO stage 1
• ds_zero2.json - DeepSpeed ZeRO stage 2
• ppo_config.json - RL training config for LLMs (future use)

---

🔗 References

• ViSearch-R1 Paper: arXiv:2503.09516
• DeepRetrieval Paper: arXiv:2503.00223

---

🤝 Contributing

This project is developed for the VLSP 2025 competition.

• The current competition pipeline does not yet integrate LLM-based retrieval, but the system supports training LLMs for query rewriting, retrieval, and reranking.
• Future pipeline versions can leverage these LLM components to improve retrieval accuracy and reasoning.
• For questions, issues, or contributions, follow competition guidelines and submit issues or pull requests.