A unified benchmark for document-centric multimodal retrieval-augmented generation (MM-RAG). This project provides the first large-scale, realistic benchmark for MM-RAG built from 70k real-world PDF pages across eight domains, with tools for document tagging, dataset synthesis, baseline implementations, and evaluation frameworks.

• Overview
• Key Findings
  • Performance Insights
  • Benchmark Characteristics
  • Analysis Capabilities
• Project Structure
• Main Components
  • Document Tagging
  • Dataset Synthesis
  • Baseline Implementations
  • Evaluation Framework
• Quick Start
  • Prerequisites
  • Document Tagging
  • Dataset Synthesis
  • Run Baselines
  • Evaluation
• Configuration
  • Environment Setup
  • Data Preparation
• Usage Examples
  • Document Processing
  • QA Generation
  • Evaluation
• Contributing
  • Development Setup
• Citation
• License
• Support
• Acknowledgments

UNIDOC-BENCH is designed to evaluate and benchmark multimodal document understanding systems across various domains including healthcare, finance, legal, education, and more. The benchmark extracts and links evidence from text, tables, and figures, then generates 1,600 multimodal QA pairs spanning factual retrieval, comparison, summarization, and logical reasoning queries. It supports apples-to-apples comparison across four paradigms: (1) text-only, (2) image-only, (3) multimodal text-image fusion, and (4) multimodal joint retrieval.

Based on the comprehensive evaluation in the UNIDOC-BENCH paper, several important insights emerge:

• Multimodal text-image fusion RAG systems consistently outperform both unimodal and jointly multimodal embedding-based retrieval
• Neither text nor images alone are sufficient for optimal document understanding
• Current multimodal embeddings remain inadequate for complex document-centric tasks

• 70,000 real-world PDF pages across 8 diverse domains
• 1,600 multimodal QA pairs with 20% expert validation
• Four query types: factual retrieval, comparison, summarization, and logical reasoning
• Unified evaluation protocol with standardized candidate pools, prompts, and metrics

• When and how visual context complements textual evidence
• Systematic failure mode identification
• Actionable guidance for developing robust MM-RAG pipelines

UNIDOC-BENCH/
├── src/
│   ├── vllm_tagging.py          # Document tagging and classification
│   ├── qa_synthesize/           # QA dataset synthesis pipeline
│   ├── baselines/               # Baseline RAG implementations
│   ├── evaluation/              # Evaluation metrics and tools
│   └── tagging_prompts/         # Prompts for document tagging
├── data/
│   └── QA/filtered/             # Filtered QA datasets by domain
└── docs/                        # Documentation and examples

Automated document tagging and classification system that processes PDF documents and extracts:

• Domain classification (healthcare, finance, legal, etc.)
• Language detection
• Date extraction
• Modality identification (text, images, tables)
• Format analysis

Key Features:

• Multi-modal document processing using VLLM
• Layout analysis with LayoutParser
• Support for both single PDFs and batch processing
• Integration with Qwen-VL models

Complete pipeline for generating high-quality question-answer pairs from document collections:

Pipeline Steps:

1. Knowledge Graph Creation (1_kg_create.py) - Build knowledge graphs from documents
2. QA Synthesis (2_qa_synthesize.py) - Generate QA pairs using RAGAS framework
3. Quality Filtering (3_filter_qa.py) - Filter out low-quality pairs
4. Similarity Filtering (4_filter_similarities.py) - Remove duplicate/similar questions
5. Balancing (5_balance.py) - Balance question types and difficulty
6. Rewriting (6_rewriting.py) - Improve question quality and diversity

Supported Domains:

• Healthcare
• Finance
• Legal
• Education
• Energy
• Construction
• Commerce & Manufacturing
• CRM

Dataset Statistics:

• 70,000 real-world PDF pages
• 1,600 multimodal QA pairs
• 20% validated by multiple annotators and expert adjudication
• Covers factual retrieval, comparison, summarization, and logical reasoning queries

Multiple RAG baseline implementations supporting four paradigms:

• Text-only RAG (simple_rag.py) - Traditional text-based retrieval
• Image-only RAG (image_rag.py) - Visual document retrieval
• Multimodal Text-Image Fusion (image_text_rag.py) - Combined text and image retrieval
• Multimodal Joint Retrieval - Joint embedding-based retrieval
• Gemini Integration (gemini_call.py) - Google Gemini model integration
• Voyage Embeddings (image_rag_voyage.py) - Voyage AI embeddings

Supported Models:

• GPT-4, GPT-3.5
• Gemini Pro
• Qwen-VL
• Various embedding models (OpenAI, Voyage, etc.)

Comprehensive evaluation suite using RAGAS metrics:

• Correctness Evaluation (correctness.py) - Answer accuracy assessment
• Retrieval Evaluation (retrieval_eval.py) - Retrieval quality metrics
• RAGAS Integration (evaluation_ragas.py) - Full RAGAS evaluation pipeline
• Analysis Tools (analysis.py) - Results analysis and visualization

Evaluation Metrics:

• Answer Correctness
• Context Precision
• Context Recall
• Faithfulness
• Answer Relevancy

# Install dependencies
pip install -r requirements.txt

# Set up environment variables
export OPENAI_API_KEY="your-openai-key"
export GOOGLE_API_KEY="your-google-key"  # For Gemini

python src/vllm_tagging.py \
    --input_path /path/to/documents \
    --output_path /path/to/tagged_output \
    --model_name "Qwen/Qwen-VL-Chat"

# Run the complete synthesis pipeline
bash src/qa_synthesize/scripts/create_dataset.sh

# Text RAG baseline
bash src/baselines/scripts/text_rag.sh

# Multimodal RAG baseline
bash src/baselines/scripts/img_text_rag.sh

# Run end-to-end evaluation
bash src/evaluation/scripts/eval_e2e.sh

1. Conda Environment:

conda create -n unidoc-bench python=3.9
conda activate unidoc-bench
pip install -r requirements.txt

1. API Keys:

export OPENAI_API_KEY="your-key"
export GOOGLE_API_KEY="your-key"
export VOYAGE_API_KEY="your-key"  # Optional

⚠️ Note: The full dataset (70k PDF pages) is too large for GitHub. Please download it separately.

1. Download the Dataset:

   • 📁 Hugging Face Dataset: Salesforce/UniDoc-Bench - Complete dataset with all PDFs and QA pairs

   The dataset includes compressed PDF files for each domain that can be used for RAG database construction:

   • commerce_manufacturing_pdfs.tar.gz (529 MB)
   • construction_pdfs.tar.gz (885 MB)
   • crm_pdfs.tar.gz (704 MB)
   • education_pdfs.tar.gz (816 MB)
   • energy_pdfs.tar.gz (708 MB)
   • finance_pdfs.tar.gz (360 MB)
   • healthcare_pdfs.tar.gz (1.19 GB)
   • legal_pdfs.tar.gz (541 MB)

   📊 Compressed File Sizes

1. Document Structure:

data/
├── final_database/
│   ├── healthcare/     # PDFs (download separately)
│   └── ...
└── QA/
    └── filtered/               # Processed QA datasets (included in repo)

1. Setup Data Directory:

# Create data directory structure
mkdir -p data/final_database
mkdir -p data/QA/filtered

# Download dataset from Hugging Face
from huggingface_hub import hf_hub_download
import tarfile

# Download and extract PDFs for each domain
domains = ["healthcare", "education", "construction", "crm", "energy", "finance", "commerce_manufacturing", "legal"]

for domain in domains:
    # Download PDF archive
    archive_path = hf_hub_download(
        repo_id="Salesforce/UniDoc-Bench",
        filename=f"{domain}_pdfs.tar.gz",
        repo_type="dataset"
    )
    
    # Extract to data directory
    with tarfile.open(archive_path, 'r:gz') as tar:
        tar.extractall("data/final_database/")

# The QA datasets are already included in the repository

from src.vllm_tagging import process_documents

# Process a single PDF
results = process_documents(
    input_path="document.pdf",
    model_name="Qwen/Qwen-VL-Chat"
)

# Process a folder of documents
results = process_documents(
    input_path="documents/",
    model_name="Qwen/Qwen-VL-Chat",
    batch_size=10
)

from src.qa_synthesize.kg_create import create_knowledge_graph

# Create knowledge graph
kg = create_knowledge_graph(
    name_str="healthcare",
    database_path="data/final_database/healthcare_database"
)

from src.evaluation.evaluation_ragas import evaluate_ragas

# Evaluate a RAG system
results = evaluate_ragas(
    input_file="results.json",
    output_file="evaluation.json",
    testsize=500
)

We welcome contributions! Please see our Contributing Guide for details.

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Add tests if applicable
5. Submit a pull request

If you use UNIDOC-BENCH in your research, please cite:

@article{peng2025unidoc,
  title={UNIDOC-BENCH: A Unified Benchmark for Document-Centric Multimodal RAG},
  author={Peng, Xiangyu and Qin, Can and Chen, Zeyuan and Xu, Ran and Xiong, Caiming and Wu, Chien-Sheng},
  journal={arXiv preprint arXiv:2510.03663},
  year={2025}
}

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC-BY-NC 4.0) - see the LICENSE file for details.

🚫 This dataset was generated using GPT-4.1 and should not be used to develop models that compete with OpenAI.

For questions and support:

• Open an issue on GitHub
• Contact: [becky.peng@salesforce.com]

• Built on top of RAGAS framework
• Uses LlamaIndex for RAG implementations
• Integrates with various LLM providers (OpenAI, Google, etc.)