okayish handles the checks that exact rules can’t: is this ticket an outage, which country is “Alemania”, is $ dollars or pesos. You declare a Pydantic-style schema, pass in messy text, and a fast decision model (Jev) chooses among the allowed values. Every field comes back typed, with a confidence score, and anything under its threshold is marked for review.
How it works
Literal and Enum fields become a Jev choice question and bool fields a yes/no probability. Exact label matches skip the model call, and with no backend configured unresolved values are simply flagged. parse_many dedupes inputs and runs calls concurrently, and parse_frame adds value, confidence and needs_review columns to a DataFrame.
Add-ons
Optional modules clean countries (ISO 3166), currencies (ISO 4217), units and job seniority. They try a local exact or fuzzy match first and call the model only for the remainder, which keeps cost and latency down. Ambiguous cases, like the $ symbol, resolve to a best guess with low confidence so they get reviewed.
Status
Early stage. The tests run against a mock backend, so no network or API key is needed, and the roadmap covers a Polars helper, caching, local backends and benchmarks. It is not meant for decisions about people.