Sift emoji out of the pile by description

01 · Role in the system

What Jev does here

Keyword search cannot answer things a magnet could attract (no emoji name mentions metal) or i need to lose weight (a goal, not a description). The request sends the query once as state and asks one independent score question per emoji - a position on a four-level scale from unrelated to an obvious answer - so all answers come in parallel, the payload stays the same size whatever the user typed, and the user’s text never enters the instructions, a stronger boundary than escaping it. Ranking rather than yes/no per emoji is the load-bearing choice: things that make noise is true of almost any object, so a fixed bar returned 27 matches for one query and 2 for another; ranking asks which emoji answer best. Code owns what happens next - rank by score, drop below a floor, take the top few - and the gateway key stays server-side.

Ranking instead of a threshold, and why it mattered

The load-bearing design decision in this little picker is rank-dont-filter. An early threshold version returned 27 matches for one query and 2 for another, because things that make noise is arguably true of almost every object — a fixed bar over fuzzy judgments is unstable. Ranking asks which emoji answer the query best, and code applies the floor and the top-few cut afterwards, where the policy can be tuned without touching a prompt.

The request shape is the other quiet lesson: the user's query travels once as shared state, one independent score question per emoji rides in the same parallel request, and the payload stays the same size whatever the user typed. The user's text never enters the instructions at all — a stronger boundary than escaping it would be.

A four-level scale with declared owners

Scores sit on a four-level scale from unrelated to an obvious answer, and the record is explicit that both the scale and the floor are code's choices, not model properties. That attribution matters when copying the pattern: the interesting design work is in the scale and the cut, and the model supplies calibrated positions on it.

The honest gaps: the README's hard-query examples were not independently rerun at verification time, calls route through an AI gateway rather than the native endpoint, and there is no hosted demo — it is a local-run repository. As search-cluster evidence it demonstrates per-candidate scoring over a static candidate set, the simplest member of the reranking family in this index.

02 · Control boundary

Where Jev sits

Query lives in state once; one score question per emoji in a single parallel request; deterministic code ranks, floors, and truncates.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• A repo to run locally (Node 22+, gateway key server-side); no hosted demo at capture time.
• Calls go through Vercel AI Gateway rather than the native TypeSafe endpoint.
• The four-level scale and the floor are code’s choice; the README’s hard-query examples were not independently rerun here.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• nischal94 ↗Community · github · public · checked 2026-09-22