Why Museum Search Is Broken and What Semantic Search Fixes
Museum catalogs organize works by fields such as title, artist, and medium. Retrievals lets signed-in users search a curated NGA image set with a text description.

Imagine a painting you remember for its warm candlelight, solitary figure, and melancholy mood. You type "warm candlelight Dutch melancholy" into a museum search box.
If a catalog search matches only its metadata fields, it may miss a description whose words do not appear in the record.
The NGA's search is not unusual. Almost every major museum search works the same way: it matches against catalog fields. Title, artist, medium, accession number, date range, classification. These fields are useful if you already know what you're looking for. They are nearly useless if you don't.
The Fundamental Problem
Museum catalog search is built around the assumption that search is a lookup. You have a target: a specific Rembrandt, a particular genre, and you use the interface to retrieve it. The system is optimized for provenance research, acquisition tracking, and curatorial retrieval. It was never designed for discovery.
The consequence is a hard prerequisite: to find a work, you must already know something precise about it. Artist name, a word from the title, or a classification term the cataloger happened to use: if your mental model of the work doesn't match the catalog vocabulary, you're out.
flowchart LR
A["What you remember\n'warm light, woman reading, Dutch'"]
B["Keyword search"]
C["Catalog fields\ntitle · artist · medium · date"]
D["No match"]
A --> B
B --> C
C --> D
classDef memory fill:#f5f0e8,stroke:#a86845,color:#2c2926;
classDef system fill:#ebe5d9,stroke:#6f685f,color:#2c2926;
classDef fields fill:#fcf9f2,stroke:#d1c7b7,color:#2c2926;
classDef result fill:#fffaf0,stroke:#a86845,color:#2c2926;
class A memory;
class B system;
class C fields;
class D result;
Even when a painting is in the collection, the interface may not connect the words you remember with the terms in its catalog record.
What Semantic Search Changes
Retrievals accepts text queries from signed-in users and searches 40,368 curated NGA images. It does not accept image uploads. Quantized EmbeddingGemma 2 maps the text into the same 768-dimensional space as the artwork images. The CPU service then ranks the images by cosine similarity. The architecture post explains how the index is built and how pagination works.
This gives signed-in users a way to enter visual descriptions even when those words do not appear in catalog fields. The cosine score measures similarity between vectors; it does not confirm a work's identity or mood.
flowchart LR
A["Signed-in text query\n'warm light, woman reading, Dutch'"]
B["Quantized EmbeddingGemma 2\nnormalized 768D vector"]
C[("40,368 curated\nNGA image vectors")]
D["Top cosine scores\n64 results, then pages to 256"]
A --> B
B --> D
C --> D
classDef memory fill:#f5f0e8,stroke:#a86845,color:#2c2926;
classDef model fill:#ebe5d9,stroke:#6f685f,color:#2c2926;
classDef space fill:#fcf9f2,stroke:#2c2926,color:#2c2926,stroke-width:2px;
classDef result fill:#fffaf0,stroke:#a86845,color:#2c2926;
class A memory;
class B model;
class C space;
class D result;
The query can use ordinary words instead of catalog terms. Results still depend on the indexed images and vector scores, so they may not identify the painting you remember.
The Difference in Practice
The two systems compare different things.
| Search | What it compares |
|---|---|
| Catalog search | Query words against fields such as title, artist, medium, and date |
| Retrievals | A text query vector against normalized image vectors from the curated NGA subset |
Catalog records may not have fields for mood, light, or how a painting feels. A text query lets you describe those qualities, but the ranking does not verify that the results match your intent.
What This Doesn't Fix
Semantic search changes what you can ask, but it doesn't change what's in the collection.
The NGA open-access dataset has 68,816 object records. Retrievals indexes 40,368 curated images from that set, so it does not search every record or every NGA holding.
You can describe visual qualities such as light, composition, color, mood, or subject. Retrievals accepts text only and does not accept image uploads. Sign in before searching.
Use the results as suggestions, then check an artwork's NGA record before identifying it.
The Underlying Shift
The difference between keyword search and semantic search is a difference in what the system assumes about the user. Keyword search assumes you know the vocabulary of the domain. Semantic search assumes you know what you want to find.
For a museum collection, that shift matters. Signed-in visitors can search 40,368 curated images from the NGA's 68,816-object open dataset with a text description. This can help people who remember standing in front of a painting but not the artist's name.
Retrievals is built for that second group.