Tagging & semantic search

Tessera describes every asset for you and indexes it for search. Nothing here requires manual tagging to get started — the library is browsable and searchable on day one.

Auto-tagging

Each asset is run through structured taggers that assign tags across categories: person, clothing, content type, pose, composition, setting, location, lighting, mood, rating, and free-form tags. Each tag carries a confidence value, rendered honestly in the interface — never faked or rounded into decoration.

Tags are data you can filter on. Combine several at once with AND logic to narrow a large library quickly: person plus setting plus rating, for example.

Captions

A local vision-language model writes a natural-language caption for each asset. Captions are full-text indexed, so a keyword in a caption is a search lane of its own, alongside structured tags.

Semantic search

Tessera embeds every image into a high-dimensional vector and searches by meaning, not filename. Type a description — “golden hour on the beach,” “person in a red coat” — and the nearest matches are ranked by how close the pixels are to your query.

This is layered, not exclusive:

• Natural-language queries search the embedding space and captions.
• Visual similarity — pick any asset and find more like it from the image embedding directly.
• Facet filters — narrow by person, place, rating, or any tag, and combine them with a semantic query.

Keyboard-first workflow

Tessera is built for long curation sessions. A command bar and keyboard shortcuts are primary affordances: open the command palette, jump between assets, flag keep/reject, and move through the filmstrip without leaving the keyboard. The grid and inspector virtualize, so scrolling a large library stays instant.

What runs where

Tagging, captions, and embeddings are all computed by local models on the compute backend you chose during first-run setup. None of it leaves your machine.