Fuzzy match
An old translation offered because its source text is close, but not identical, to the new string. Useful head start, mandatory review.
What it is
A fuzzy match is a translation-memory suggestion whose source is similar to the new segment but not identical. If the old source was 'Delete this file?' and the new one is 'Delete these files?', the TMS may surface the old translation with a score so the translator can edit it instead of starting blank.
Use fuzzy matches to speed up revisions, repeated product patterns, and large migrations where much of the copy changed only slightly. The score is a ranking aid, not a universal measurement: tokenization, punctuation, placeholders, and context penalties differ between tools, so 85 percent in one TMS is not the same promise in another. Vendor pricing grids still use the Trados fuzzy bands, typically 75 to 84, 85 to 94, 95 to 99, and 100 percent, each discounted differently.
Gotcha: one changed word can reverse the meaning. Numbers, negation, product names, and ICU placeholders deserve special attention even at a high score. Keep fuzzy suggestions visibly unapproved until a translator reviews them, and never charge or report them as exact matches merely because the target text was copied forward.
Ask AI for it
Build fuzzy translation-memory suggestions with PostgreSQL pg_trgm. Normalize whitespace, preserve case and negation, replace each ICU MessageFormat argument with a typed placeholder token, and rank stored source segments with the similarity function. Return the top five above a configurable threshold with source text, target text, locale, context, and score. Penalize missing or changed placeholders, highlight the exact token differences, and mark every non-identical result as 'needs-review'; never write it directly into an approved translation.