Smart search
How plain English becomes a query string, what checks the output, and how often it gets it right.
Smart search turns a sentence into a query string. It is the least grounded thing in the app: one model call with no database access, no tools and no retrieval. What makes it usable is that the output is a query, and a query can be checked before it runs.
The path
The prompt is a syntax reference plus worked examples, and an output
contract demanding exactly one line with no prose. It carries the traps we have
watched the model fall into: that "cheap" means mana value while "budget" means
dollars, and that t:(instant OR sorcery) is invalid syntax in Scryfall and
here, despite being a thing models reliably produce.
The sanitizer strips the output down to a query when the model adds prose anyway.
Validation parses the result and checks tag names against the taggable
concepts in the database, so tag: cannot reference something that does not
exist.
One retry. A failed query goes back with its specific errors attached.
Both fallbacks favour showing you something over showing you nothing. A retried query is returned without being checked a second time, and if the sanitizer rescues nothing the raw text is used. For a search bar that is the right tradeoff, since you can see and edit what it produced. It does mean a query can reach the bar without having passed validation, and today nothing marks it as such.
How reliable it is
The translator is evaluated against a corpus covering plain requests, ambiguous phrasing and deliberate traps. Most translations are usable as written. A minority need an edit, usually because the model picked a defensible but different reading of an ambiguous request.
Two failure modes are worth knowing about. It sometimes ignores the one-line output contract and emits reasoning alongside the query, which is what the sanitizer exists to catch. And it occasionally misreads cost words, reaching for mana value where you meant dollars. Both are areas of active work.
The honest summary is that this produces a good first draft with a repair layer under it, and the draft is shown to you rather than run behind your back.
What to do about it
The translated query goes into the search bar rather than running invisibly, so you can read it before trusting the results. When it is wrong, editing the query is faster than rephrasing the sentence, and the syntax reference is short.
There is a feedback control on the results. It records what you asked, what you got, and what you expected, which is where the eval cases come from.