Search that answers what you meant

LIKE '%stressed%' finds nothing when the asset says pressing the bridge of their nose with eyes shut — which is exactly the picture you wanted. So queries are expanded through a concept map built for this catalogue's own subject matter, then scored by term frequency against inverse document frequency, so a rare word like greenhouse counts for far more than office in a library that is mostly offices.

It shows its working

Every search tells you which words it also looked for, because a search box that silently rewrites your query is a search box you cannot steer. Also looked for colleagues, table, conference, boardroom — because that is how this catalogue words it.

It admits when it has nothing

A ranked search that tops out below a real match does not pretend the near-misses are answers. It says so, and offers to make the thing instead. Relevance is a number on every result, not a mood.

Filters that narrow first

Someone who asked for a portrait does not want the best landscape instead, so kind, category, shape and colour narrow before relevance ranks. Colour is a real palette — the dominant colours of every asset are measured and filed into families when it is published.

More like this, by looks and by meaning

Related assets are found with a perceptual hash for how a picture looks and a term vector for what it is about, weighted together. Two copies of the same picture are not similar — they are the same asset twice, and the catalogue refuses to list the second.