Khuyen Tran tests a product-matching pipeline on the Abt–Buy benchmark. String similarity and embeddings retrieve plausible matches. TypeSafe’s Jev then chooses among a short list, including “none of these,” using rules about model number, capacity and variant.

Retrieve candidates, then decide

  • The reported RapidFuzz baseline makes 341 wrong merges when forced to pick one match for each of 1,081 Abt records. A Sentence Transformers nearest-neighbor baseline makes 318. Neither baseline has a review or no-match option in this comparison.
  • The pipeline retrieves the top three names using all-MiniLM-L6-v2 embeddings. Jev sees fuller product details and a Choice question that includes “none of these.” A correct match outside the shortlist cannot be recovered by the decision step.
  • At the author’s 0.7 merge cutoff, the article reports 807 correct and one wrong among 808 automatic merges. It sends 62 records to review and leaves 211 unmatched. Those are the author’s results, not independently reproduced measurements.

What the cost claim does and does not cover

The article reports about 35 seconds and $0.03 for its Jev stage. Its companion README instead says 90 seconds and $0.07, and describes 3,243 pairs sent to Jev. The inspected script makes one system_one call per Abt record with three candidates per call; the repository does not commit result rows. The two timing and cost figures cannot be reconciled from the published artifacts.

The script calculates cost from Jev input tokens. It does not include the local embedding work, review labor or error-handling costs. A threshold chosen on these examples also needs validation on records held out from threshold selection. Typed probabilities make an abstention policy possible; they do not establish calibrated correctness on a new catalog.

Verification: read the full article and inspected the companion match.py, download script and README at commit 169dd24. Did not run the benchmark, download the dataset or call the paid Jev API. The companion repository contains no saved predictions to recompute the reported counts.