Digital Collections Explorer

Made for galleries, libraries, archives, and museums

A search site for your digital collection, with or without a catalog.

Digital Collections Explorer is an open-source platform for exploratory search over digital collections with visual features. Point it at a folder or an S3 bucket, and in a few steps you have a website where visitors search in natural language or with an image, even when the collection has little or no descriptive metadata.

Your collectionPhotographsMapsDocumentsA visitor's searcha bird's-eye view of a harborindexed onceeach searchOpen modelSigLIP or CLIP, locallymost similar firstYour collectionindexed oncePhotographsMapsDocumentsA visitor's searcheach searcha bird's-eye view of a harborOpen modelSigLIP or CLIP, locallymost similar first

Who it is for

Collections that are hard to search

Manual cataloging does not scale to millions of items, so many digitized and born-digital collections lack basic descriptive metadata, or have it for some items and not others. Search here compares a visitor's words with what the images show, and reads the catalog wherever there is one.

  • Photographs nobody has cataloged

    Thousands of scans named IMG_2041.jpg. Point it at the folder and visitors can search them the same afternoon.

    A search for children playing in the snow, answered with a greeting card, a snowy Hiroshige print and a skating scene

  • Maps, plans and drawings

    Search by what is drawn rather than what the title says, and open a large sheet in deep zoom.

    A search for a bird's-eye view of a city, answered with panoramic views of St. Louis and Boston

  • Born-digital and scanned documents

    PDFs are indexed page by page and searched by what each page looks like, so handwriting, diagrams and forms are found without OCR.

    A search for a handwritten letter, answered first with handwritten pages from the Smithsonian Institution Archives

  • Objects scanned in 3D

    Visitors turn the model on the object's page. Views drawn from it are searched with the photographs, and similar shapes are a click away.

    The 3D model of Charlie Parker's alto saxophone on its page, above the object's photographs

How it works

From a collection to a search site in three steps

  1. Point it at a collection

    A folder, an S3 bucket, or a parquet manifest of images and their catalog. A profile of the collection first shows what it holds, what the site will do with it and what indexing will cost.

  2. Build the index

    An open multimodal model embeds every image on a machine you control. No image is sent to an external API.

  3. Open the search site

    Visitors search in natural language or with an image, across photographs, maps, documents and 3D models. Where a catalog exists, search weighs what the images show against what the records say, and tells the visitor how it read the query.

How search works

amelia earhart

  • Amelia Earhart 'Round The World Flight' envelope signed by Amelia Earhart, National Postal Museum
  • Token commemorating Amelia Earhart's first transatlantic flight, National Postal Museum
  • Medal presented to Amelia Earhart, National Postal Museum
  • Amelia Earhart 'Round The World Flight' envelope, National Postal Museum
A name is looked up in the records, so “amelia earhart” finds what is cataloged under her name.

two men shaking hands

  • Samuel Clemens and John T. Raymond shaking hands, a photograph, National Portrait Gallery
  • Silhouettes of David Osgood and George Parker facing each other, National Portrait Gallery
  • Silhouettes of Nathan Crocker and William Goddard facing each other, National Portrait Gallery
  • Silhouettes of Theodore McCaleb and Isaac Johnson facing each other, National Portrait Gallery
A description is matched against the pictures by an open model, such as SigLIP, SigLIP 2 or CLIP. None of these records mentions a handshake.

Showcase

Try it on a real collection

Agentic search

Searchable by AI agents

Every site comes with its own MCP server. Connect Claude, ChatGPT or any other agent that speaks the Model Context Protocol, and it can search the collection on its own: run searches, follow similar objects, and link each answer to the object's page on your site.

Turn it on and connect an agent

Get started

Three ways to start

Community

Built in the open

Help decide what comes next.

Wishlist

Ideas live in the project's Discussions. Anyone with a GitHub account can post one or vote for one. The most wanted ideas are what gets built next.

Post an ideaSee all and vote

Help without writing code

  • Try it on your own collection and tell us what the search got right and wrong. Real material from real institutions is the feedback we cannot get any other way.
  • Write up how you used it so the next archive with the same problem can find the answer.
  • Judge search quality for a kind of material you know well: maps, posters, herbarium sheets, textiles.

Cite

Using it in research? Cite the paper