Ex Regex

A Python library in the shape of the standard re module that finds, replaces, splits and extracts text by what it means. A unit proposes candidate spans and Jev judges each span against a plain-English description, separating the pattern-matching job from the meaning job.

Each proposed span becomes a yes-or-no question against your description, with a probability; up to 32 questions share one request and eight requests run at once. Every Match carries p, and you choose the threshold, with bands yes at 0.9 or above, no at 0.1 or below and review in between.

Jev never writes text and can only pick a span a finder found, so it cannot transpose a digit or return a value that was not in the input. Every match is still a verbatim slice of your text with exact offsets.

Run python -m pip install --pre ex-regex; Python 3.10 or newer, no runtime dependencies. Live calls need a key: export OPENROUTER_API_KEY for Jev through OpenRouter, or use TYPESAFE_API_KEY for TypeSafe directly.

On 87 authored scenarios with Jev 1.13 through OpenRouter: personal contact redaction hit 100% precision and recall against 42% and 44%, refund requests 97% against 53%, and total due 100% exact against 47%. The full evaluation used 114 requests and cost $0.0025.

Yes. diff audits a pattern against the meaning and prints only disagreements, prefilter narrows candidates so failing spans cost nothing, any span-returning function can serve as a unit, and a recorded decision cache can be committed and replayed offline in CI with --replay.

scan shows every span with its probability, extract returns the one span that matches, filter and rank judge many texts, classify returns a Pick with every label's probability, rate returns a Rating on an ordered scale, and every function has an async twin.