Semantic search for design assets: find by meaning, not filename

Watch real queries fail. Someone searches “launch visual” against an asset titled “Q3 hero — dark gradient”: zero shared words, zero results, and the searcher concludes the asset doesn't exist — the most expensive wrong answer a library can give, because now they rebuild it.

The gap has three recurring shapes: synonymy (visual vs. hero vs. illustration), abstraction (“moody” vs. a palette), and language — on a bilingual team, half the queries arrive in a language the titles weren't written in. No amount of naming discipline closes all three; names can't anticipate every phrasing a future searcher will try.

None of this retires keyword search. Full-text over titles and tags is instant, exact, and transparent — when the searcher knows a real word from the asset, nothing beats it. Three properties are worth demanding from the keyword layer:

• Every field indexed — titles, tags, descriptions, client and product names, so one remembered fragment is enough.
• Language-neutral indexing — a bilingual team needs both languages treated as first-class, not English stemming applied to French words.
• Ranking that respects fields — a title match outranks a description match, so precise queries stay precise.

The mechanism, minus the mystique: every asset's textual identity — title, tags, description — is converted into an embedding, a long list of numbers encoding what the text is about. Meaning lands in position: texts about similar things get nearby numbers. A query is embedded the same way, and results are ranked by distance. No words need to match; the aboutness does.

“moody launch visual” ≈ “Q3 hero — dark gradient” · similarity 0.87

That single property dissolves all three gap shapes at once: synonyms sit close together, abstractions sit near their concretes, and translations sit near their originals — which is why the French query finds the English-titled asset without anyone maintaining a glossary.

Keyword wins on known-item search (“the SOC 2 one-pager”), on filters (“all banners for product X” is a tag query), and whenever precision matters more than recall. It's also free of surprises — you can see why a result matched.

Semantic wins on descriptions from memory, on cross-language queries, and on exploration — “something warm for the newsletter header” is a browse expressed as a sentence.

The third door is an example. Sometimes the query is an asset: here's one, show me its neighbors. “Find similar” assembles variant sets in seconds and doubles as hygiene — near-identical neighbors at the top of the list are usually a supersession that escaped.

In practice the best front door runs both at once — a hybrid where exact matches surface first and meaning fills in behind — so the searcher never chooses a mode; they just type.

Semantic search is downstream of metadata. The embedding can only encode the meaning your library expressed — which makes the unglamorous habits from the tagging workflow and the import gesture the actual AI strategy: descriptive titles, three-to-six honest tags, types set at import. AI-drafted tags close the loop neatly — the scanner describes the visual, the description enriches the embedding, and search gets sharper with every reviewed import.

Expectations, honestly set: on a fifty-asset library, keyword search alone feels fine. The gap opens with scale and staff turnover — at five hundred assets, when the person who named everything has left, meaning-based retrieval is the difference between a library and an archive.