MONET Retrieval

MONET Retrieval

A Gradio app from jasperai that finds the nearest matches in the MONET dataset from an uploaded image or a caption keyword, using DINOv2, CLIP or SSCD embeddings.

Made by jasperai on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: MONET Retrieval is a self-contained Gradio app from jasperai that searches the MONET dataset by nearest neighbour. You can upload an image and get its closest matches, or type keywords and search captions. It queries precomputed FAISS indices, and it runs on zero-a10g hardware at Hugging Face.

Who it suits: Researchers who work with the MONET dataset and want to look up similar items, and developers who want a worked example of an embedding search app in Gradio.

Industry | Who uses it | How they use it | Example
Research | Dataset researchers | Upload an image and look at its nearest neighbours | Finding items in the dataset that resemble a query picture
Machine learning | Embedding engineers | Compare how DINOv2, CLIP and SSCD rank the same query | Seeing which embedder returns the closest matches
Software development | Gradio developers | Read app.py and embedders.py | Studying how three embedders are loaded and queried

Skip it if you want a general image search over the web, since it only searches the MONET dataset.

Where to use it: MONET Retrieval runs in these places.

• A web browser: open the Space's page on Hugging Face
• Python 3.12.12 on Windows, macOS or Linux: run a copy of the Space

Use case | Who | What they get back
Search by image | Researchers | The k nearest neighbours of an uploaded image, as thumbnails
Search by caption | Researchers | Keyword matches found through CLIP text embeddings
Read a record | Anyone | Full record metadata, shown when you click a thumbnail

How to set it up: Online use needs only a browser. A local copy needs more, because the app expects its index files in a set layout.

• A web browser for the online Space
• Python 3.12.12 on Windows, macOS or Linux for a local copy
• The Python packages listed in requirements.txt
• The index files, shipped in a storage bucket named jasperai/monet-retrieval-storage

1. Install the Hugging Face hub library with pip.
2. Download the Space's files into a folder called monet-retrieval.
3. Move into that folder.
4. Install Gradio at the version the Space uses.
5. Install the packages listed in requirements.txt.
6. Make the index files available, then start the app with Python.

pip install -U huggingface_hub
hf download jasperai/monet-retrieval --repo-type space --local-dir monet-retrieval
cd monet-retrieval
pip install gradio==6.13.0
pip install -r requirements.txt
python app.py

The app reads optional environment variables. These are the main ones.

Name | Default | What it does
MONET_BUCKET_MOUNT | /data | Where the storage bucket is mounted
MONET_DATA_ROOT | next to app.py | The folder the bucket files are copied into
MONET_MODELS | dinov2,clip,sscd | Which embedders to load
MONET_SKIP_SYNC | off | Set to 1 to read index files straight from the bucket instead of copying them
MONET_HF_REPO | jasperai/monet | The dataset that thumbnails are fetched from
HF_TOKEN | none | Hub access for private datasets
MONET_DOWNLOAD_WORKERS | 16 | How many thumbnail fetches run in parallel
MONET_SYNC_WORKERS | 4 | How many copies run in parallel during the bucket sync
MONET_PARQUET_CACHE_BUDGET_GB | 8 | Size cap for the fallback cache, where 0 turns caching off
MONET_PARQUET_CACHE_FREE_FLOOR_GB | 2 | Free disk level below which the cache is cleared

How to use it: Online, open the Space's page and pick a search mode. On your own computer, start the app from the folder with the command below.

python app.py

Tool | What it does
Search by image | Embeds an uploaded image with DINOv2, CLIP or SSCD and returns its nearest neighbours
Search by caption | Runs a keyword search through CLIP text embeddings
Thumbnail click | Shows the full record metadata for that item

Example: You upload a photo and ask for its nearest neighbours with one of the three embedders. The app returns matching thumbnails, and clicking one shows that item's full record metadata.

For newcomers and experts: If you are new, open the Space's page and upload a picture. An embedding turns an image into numbers, and the app finds the entries whose numbers are closest.

If you know the field, the app preloads its three embedders at launch, so even the first query skips model loading. Thumbnails are range-read from the upstream parquet files on the Hub, so no full downloads are needed. A shared alignment table maps rows across the three models, which keeps the app from loading the large hash files. Skipping the bucket sync is recommended on Hugging Face Spaces, where the disk is too small for more than 24 GB of index files.

Open the Space or run app.py
Upload an image or type keywords
The embedder turns the query into numbers
FAISS finds the nearest neighbours
Thumbnails appear with their metadata

Good to know: It searches only the MONET dataset, and it needs its index files to work.

• Version: Gradio 6.13.0 and Python 3.12.12
• Embedders: DINOv2, CLIP and SSCD, where CLIP alone is named for text search
• Size: 8 files in the repository, plus the separate index bucket
• License: cc-by-4
• Cost to run: none named, but the index files run past 24 GB
• Data it reaches: the MONET dataset files on the Hugging Face Hub, the index bucket and the SSCD checkpoint, which is fetched on first use
• Not for: no forbidden uses are named

On Hugging Face, 13 November 2026

• Likes: 20
• Built with: gradio
• License: cc-by-4.0

Links

• Space: https://huggingface.co/spaces/jasperai/monet-retrieval
• Files: https://huggingface.co/spaces/jasperai/monet-retrieval/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.