For teams

11MB binary. 23MB embedding model. Runs on CPU, no GPU. Queries in ~50ms, local — no per-query cost, no data leaving your servers. Catalyst generation is the only cloud call: cents per user per refresh, through your own API key.

For deployment details and cost comparisons, see the infrastructure page.

Designing your content for taste

Your users save, explore, choose, return, reject. Those actions are already a corpus. The question is how to represent them so Enzyme can read preference from the accumulation.

Curation app
Users accumulate Saved articles, bookmarked images, collected links — 800+ items over a year
Text representation Link title, source, description, tags from the collection it was saved to
Structure available Collections as folders, favorites as tags, save timestamps
Agent sees Which saves cluster by week. What the user flagged vs. what they quietly accumulated. Where collections overlap.
Example catalyst "Three clusters of saves appeared in the same week — what were you circling?"
What taste unlocks Surface why two users' collections resonate — not shared items, but shared patterns of saving
Design tool
Users accumulate Exploration sessions — options shown, choices made, variations rejected
Text representation The delta: what was offered vs. what was chosen. "Shown 30 hover states; gravitated toward spring-damped and asymmetric-ease; rejected all options with >2 accent colors"
Structure available Projects as folders, agent-applied tags like #returned-to or #rejected, session timestamps
Agent sees Preference patterns across sessions. Which constraints are consistent vs. which shift. Where the user deliberated longest.
Example catalyst "You chose the constrained option four sessions in a row — is that a preference or a phase?"
What taste unlocks Generate options that push 20% past the user's current edge — work within preferences while stretching them
Meeting / voice app
Users accumulate Transcripts, post-call reflections, decision logs — 14 months of conversations
Text representation Full transcript text, speaker attribution, any user-added notes after the meeting
Structure available Folders by team or project, tags for decision type, wikilinks to people and projects
Agent sees Which topics recur across months. Which decisions got reopened. Which people keep getting referenced in unrelated contexts.
Example catalyst "The same concern surfaced in three unrelated conversations this month. The language changed each time — what shifted?"
What taste unlocks Show users the through-lines across their conversations — not summaries, but what their meetings have been about
Reading / highlights app
Users accumulate Book highlights, article annotations, margin notes — years of reading
Text representation Highlighted passage, source title and author, any user annotation or note
Structure available Source as folder or link, user-applied tags, highlight timestamps
Agent sees Which ideas recur across different authors. Where a highlight from last year rhymes with one from this week. Annotation density as a signal of depth.
Example catalyst "You highlighted the same idea in three different books over two years. Each author framed it differently — which framing stuck?"
What taste unlocks Surface the intellectual threads running through someone's reading — the curriculum they didn't know they were building

Two things to get right

The text representation. Enzyme processes text. For products where the raw material is images, audio, or interaction data, the text you generate per artifact determines what catalysts form. An image described as “black and white photography, runway” produces catalysts about visual properties. The same image described as “deconstructed tailoring, mid-career collection, linen” produces catalysts about taste. The representation is the thing to iterate on. The pipeline stays the same.

Emergent structure over imposed taxonomy. The best structure comes from your product’s existing UX. A collection is a folder. A favorite is a tag. A project is a folder. Saving something next to something else is an implicit link. Map the organizing gestures your users already perform onto the primitives Enzyme reads — don’t ask them to categorize. Tags the agent applies (like #returned-to or #considered-and-rejected) are especially valuable: behavioral structure without user friction.

The engine is the same one that runs over notes workspaces. The pipeline doesn’t change. What changes is the input and the character of the catalysts that form.