Official repository for VaseMuseum, a multimodal agent framework for trustworthy interaction with ancient Greek pottery in virtual museum environments.
VaseMuseum: Digital Intelligent Museum for Ancient Greek Pottery
Jiazi Wang*, Nonghai Zhang*, Qiushi Xie*, Zeyu Zhang*†, Yufeng Chen, Yang Zhao, Ling Shao, Hao Tang#
*Equal contribution. †Project lead. #Corresponding author.
Paper | Project Page | Demo
Digital museums are becoming increasingly important for cultural heritage preservation, education, and public engagement. While modern Vision-Language Models (VLMs) can describe visual content effectively, they often struggle when answering specialized cultural-heritage questions that require reliable historical knowledge and evidence-grounded reasoning.
VaseMuseum addresses this challenge by integrating:
- 🏺 Ancient Greek pottery understanding
- 🖼️ 2D image and 3D artifact perception
- 🌐 External knowledge retrieval
- 🔍 Evidence verification and source control
- 🤖 Reliability-aware multimodal reasoning
At the core of VaseMuseum is VaseAgent, a multimodal reasoning agent that combines visual understanding, knowledge retrieval, and response calibration to provide trustworthy museum assistance.
- Interactive digital exhibition space
- Exploration of pottery collections
- Natural-language interaction with exhibits
- Support for both image-based and 3D artifact browsing
A multimodal cultural-heritage assistant capable of:
- Visual understanding of pottery artifacts
- Shape and iconography recognition
- Historical and archaeological reasoning
- External knowledge retrieval
- Evidence-grounded answer generation
Unlike standard retrieval-augmented systems, VaseMuseum introduces:
-
Source Control
- Link validation
- Source quality assessment
- Diversity-aware evidence selection
-
Response Control
- Claim-evidence verification
- Uncertainty calibration
- Hallucination reduction
A lightweight inference-time selection strategy improves:
- Citation validity
- Evidence support
- Neutrality under ambiguity
- Response reliability
without modifying the underlying VLM.
- Project page release
- Demo release
- Dataset release
- Open-source evaluation framework
- Additional museum collections support
VaseMuseum/
├── deploy/ # Virtual museum frontend
│ ├── DigitalExhibition/
│ ├── css/
│ ├── js/
│ └── index.html
│
├── dataset/ # Museum datasets and metadata
│ └── data/
│
├── retriever/ # Knowledge retrieval pipeline
│ ├── build_corpus.py
│ ├── pipeline.py
│ ├── local_llm.py
│ └── caption.py
│
├── vase-agent/ # VaseAgent implementation
│ ├── core/
│ ├── tools/
│ ├── metrics/
│ ├── main.py
│ └── agent_run.py
│
├── vllm_run/ # VLM serving scripts
│ ├── start_vllm_api.sh
│ └── call_vllm_api.sh
│
└── README.md
VaseMuseum consists of four major components:
Users can:
- Browse exhibits
- Inspect artifact details
- Explore 3D objects
- Ask natural-language questions
The VLM extracts:
- Vessel morphology
- Decorative patterns
- Painting techniques
- Iconographic elements
- Scene composition
When visual information is insufficient, VaseAgent:
- Searches authoritative sources
- Collects supporting evidence
- Aggregates museum and scholarly information
The system verifies:
- Evidence quality
- Source validity
- Claim support
- Response confidence
before returning answers.
This project uses uv to manage Python dependencies. All packages are declared in vase-agent/pyproject.toml (Python ≥ 3.11).
# Install uv
curl -LsSf https://astral.sh/uv/install.sh | sh
# Optional: use Tsinghua PyPI mirror for faster installs in China
export UV_INDEX_URL="https://pypi.tuna.tsinghua.edu.cn/simple"
cd vase-agent
# Create a virtual environment and install locked dependencies
uv sync
# Configure API keys and model endpoints
cp .env.example .env # then edit vase-agent/.envTo run commands inside the project environment:
uv run python main.pyOr activate the virtual environment manually:
source .venv/bin/activatecd DigitalExhibition
python -m http.server 8000Then open:
http://localhost:8000
in your browser.
Direct Inference
cd vase-agent
uv run bash infer.shExperience Accumulation
uv run bash experience.shcd retriever
python build_corpus.pyRun retrieval:
python cli.pyThe framework supports evaluation of:
- Answer Accuracy
- Groundedness
- Hallucination Rate
- Citation Validity
- Neutrality Under Ambiguity
Example:
cd vase-agent
uv run python -m metrics.llm_judge \
--input runs/task/predictions_all.jsonl \
--output runs/task/judged_per_sample.jsonl \
--aggregate-out runs/task/metrics_summary.json \
--workers 4- Digital museums
- Artifact interpretation
- Collection management
- Interactive learning systems
- Virtual exhibition guides
- Historical storytelling
- Archaeological analysis
- Iconography studies
- Cross-collection retrieval
- Museum assistants
- Exhibit Q&A systems
- Online cultural experiences
- Multi-museum integration
- Additional artifact categories
- Stronger multimodal reasoning
- Interactive agent planning
- Multilingual support
@article{wang2026vasemuseum,
title={VaseMuseum: Digital Intelligent Museum for Ancient Greek Pottery},
author={Jiazi Wang and Nonghai Zhang and Qiushi Xie and Zeyu Zhang and Yufeng Chen and Yang Zhao and Ling Shao and Hao Tang},
journal={},
year={2026}
}We thank the open-source communities and cultural-heritage institutions that support digital preservation, multimodal research, and public access to historical collections.
For questions or collaborations, please open an issue or contact the authors.