HEAVEN: Hybrid-Vector Retrieval for Visually Rich Documents

Official Repository for our paper "Hybrid-Vector Retrieval for Visually Rich Documents: Combining Single-Vector Efficiency and Multi-Vector Accuracy"

🔥News

• [2026-04] HEAVEN is accepted to ACL 2026 Findings🎉
• [2025-11] ViMDoc is now available on Hugging Face🤗

ViMDoc Benchmark

ViMDoc (Visually-rich Long Multi-Document Retrieval Benchmark) for evaluating visual document retrieval under both multi-document and long-document settings.

from datasets import load_dataset
dataset = load_dataset("kaistdata/ViMDoc", split="ViMDoc")

Format

Sample datasets are available in benchmark/{ViMDoc,OpenDocVQA,ViDoSeek,M3DocVQA}. Each contains sample_query.json with queries and ground truth document IDs:

{
    "id": "<query_id>",
    "query": "<query_text>",
    "doc_ids": ["<document_id>"]
}

Sample document pages are stored in sample_pages/.

Note: Full datasets for other benchmarks are available from their original sources: OpenDocVQA | ViDoSeek | M3DocVQA

Indexing

(1) Encoding (Query/Document)

cd indexing/encode

# Visusal encoder
python encoder.py --encoder_type dse --folder ViMDoc
python encoder.py --encoder_type colqwen25 --folder ViMDoc

# Textual encoder
python ocr.py --device 0 --folder ViMDoc
python encoder.py --encoder_type nvembedv2 --folder ViMDoc
python encoder.py --encoder_type bge_m3_multi --folder ViMDoc

Available Encoders

(2) VS-Page Construction

cd indexing/vs-page

# Step 1: Document Layout Analysis
python DLA.py --dataset ViMDoc --device 0

# Step 2: Assemble & VS-page Encoding
python assemble.py \
    --dataset ViMDoc \
    --encoder_type dse \
    --reduction_factor 15 \
    --device 0

Retrieval - HEAVEN

Run the complete HEAVEN pipeline (Stage 1 + Stage 2):

cd retrieval/heaven

python heaven.py \
    --folder ViMDoc \
    --stage1_model dse \
    --stage2_model colqwen25 \
    --device 0 \
    --preprocess

Stage 1 Only :

python stage1.py --folder ViMDoc --model dse --alpha 0.1 --filter_ratio 0.5

Stage 2 Only :

# Preprocess queries first
python preprocess.py --folder ViMDoc --model colqwen25

# Run Stage 2
python stage2.py --folder ViMDoc --model colqwen25 --stage1_model dse --k 200 --filter_ratio 0.25

Structure

HEAVEN/
│
├── benchmark/                    
│   ├── ViMDoc/                  
│   ├── OpenDocVQA/            
│   ├── ViDoSeek/                
│   └── M3DocVQA/
│       
├── indexing/                      
│   ├── encode/                  
│   └── vs-page/
│               
├── retrieval/                    
│   ├── baeline/                   
│   └── heaven/
│                
└── run.sh              

Citation

@article{kim2025hybrid,
  title={Hybrid-Vector Retrieval for Visually Rich Documents: Combining Single-Vector Efficiency and Multi-Vector Accuracy},
  author={Kim, Juyeon and Lee, Geon and Choi, Dongwon and Kim, Taeuk and Shin, Kijung},
  journal={arXiv preprint arXiv:2510.22215},
  year={2025}
}