Tags are a shared vocabulary applied privately
A Tag row is shared across every reader. The ShowTag row that applies it belongs to one
person. Two readers can agree that “slow burn” means something without either of them seeing the
other’s shelf, and matching is on the slug so that “Slow Burn” and “slow burn” cannot become two
tags.
Context
Genres come from TMDb and describe what a show is. Nothing described what it was like to watch, which is the judgement a viewer actually trades with other viewers.
The models had existed since the first schema and held zero rows. personalization.py was already
reading ShowTag on two code paths, weighting tag affinity by relevance, so the recommender had a
live input that could never contain anything.
Three shapes were possible.
Fully private tags, one vocabulary per reader, is the simplest and the least useful. Everyone invents their own spelling of the same idea, no two readers’ tags are comparable, and the recommender learns from a vocabulary of one.
Fully public tags, where applying a tag changes what everyone sees, makes a show’s tag list a thing to argue about. It also makes one bad actor’s vocabulary everyone’s problem, and it is the failure mode that turns tagging systems into moderation systems.
Shared vocabulary, private application is what MovieLens calls a genome. The word is common property; who applied it to what is not.
Decision
Tag holds the word. ShowTag holds one person applying one word to one show, with a relevance
weight, unique on (user, show, tag).
Matching is on the slug, not the name. slugify("Slow Burn") and slugify("slow burn") are
both slow-burn, so the second one reuses the first tag rather than creating a rival. Letting both
exist would halve the signal learned_tag reads, and the split would be invisible in the UI.
Suggestions come in three tiers, and the order carries the whole idea. Tags other readers put on this show rank first, ordered by how many readers chose each, because that is the closest thing to a second opinion the catalog can offer. The vocabulary in general use comes next, by frequency. Everything else is alphabetical.
Only names cross the boundary. The suggestion strip shows what a show has been called. It never shows who called it that.
Tagging saves in place, following ADR-10. The first version
redirected to #tags and threw the reader a third of the way down the page, which is the exact
behavior ADR-10 was written to eliminate.
Tags are searchable with tag: and ranked alongside genre and network in
ADR-12, because a tag is the same kind of claim about a show.
After Action Review
Written on the same day as catalog search, and the two fit together better than either was designed to. Search gave tags somewhere to be useful the moment they existed.
The cold start is worse than ratings. A rating costs one click; a tag costs a decision about what a show was like. Until several readers have tagged the same show, the suggestion strip is empty and the first tier does nothing, which is the tier carrying the design.
ShowTag.relevance is written as 1.0 for every row and nothing sets it otherwise. The field
exists because personalization.py already weights by it. Whether a reader should ever set it, or
whether it should be derived from how many people applied the same tag, is undecided and deliberately
so.
The Recommendation model was deleted in the same commit. Nothing wrote it and nothing read it:
ADR-07 had materialized SimilarShow instead, leaving a table
whose reason field was never going to hold anything.