Community · E2 · artifact verified
Rerank Postgres product search with two questions
Product search inside PostgreSQL - typos, barcodes, typeahead, facet counts, all in SQL - with an optional second stage asking Jev two questions per search to rerank the shortlist; the README is itself the report, three failed versions included.
01 · Role in the system
What Jev does here
Stage one is pure SQL: trigram and prefix matching that survives typos and barcodes, with facet counts. Stage two is the experiment: the top candidates go to Jev as two typed questions per search, and the calibrated answers rerank the shortlist before rendering. The README doubles as an honest lab report - it measures how much the second stage actually helps, documents three earlier versions that failed and the specific fixes, and archives the prompts. A rerank skill, a dedicated Jev client module, and tests over the judge path ship in the repo.
02 · Control boundary
Where Jev sits
Jev as the two-question reranker above SQL product search: deterministic trigram and prefix matching produces the shortlist, typed questions rerank it with calibrated confidence, and failed iterations are documented alongside the working prompts.
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
- Second-stage gain is author-measured on the product catalog in the repo.
- Two calls per search add latency and cost over pure SQL; thresholds are the author's.
- 2 stars; single-maintainer e-commerce scope.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- alexforman1 ↗Community · github · public · checked 2026-10-08