Skip to main content
This prerelease replaces the withdrawn 0.4.5-beta.2 and builds on beta.1. Use 0.4.5-beta.3 for Cargo and npm, or 0.4.5b3 for Python. npm packages remain on the beta dist-tag; the stable latest release remains 0.4.4.

​
Faster scanning

Text scans are 2.5–5× faster in reported measurements, with identical existing findings. The speedup depends on the input and enabled detectors.
  • Search required prefixes and separators before validating API keys, credential URIs, JWTs, UUIDs, dates, and IP addresses. Digit detectors skip text without ASCII digits, and card candidates use a fixed digit buffer.
  • Remove three quadratic slowdowns: repeated parsing of long IPv6 colon runs, repeated work during token restoration, and grouping structured tokens by leaf.
  • Convert ASCII offsets directly, clone only selected findings, and reuse selected structured leaves during transformation.
  • Release Python’s GIL while the Rust text scanner runs, allowing other Python threads to keep working.

​
Cards followed by CVVs or expiry dates

A card number followed by a CVV or expiry is now detected. For example, 4111 1111 1111 1111 123 and 4111 1111 1111 1111 12/25 detect the leading card without including the trailing digits in its span. This changes redacted output: it hides cards that 0.4.4 missed in these contexts. No 0.4.4 finding is removed. Existing whole-number card checks run first, and issuer, Luhn, grouping, and decimal guards still apply.

​
Model session routing

A model session’s scan(text, config) routes each requested entity type to rules or the model. In Node.js, pass the scan configuration under scan(text, {config}). Omitting configuration equals {}: all enabled rule detectors run, together with the model’s other labels. Rules handle supported entity types even when the model also labels them. A rules-only request never uses or reserves the model worker and does not return model_busy; closed or failed sessions still reject requests. Use scan_model in Rust or Python, or scanModel in Node.js, for raw model output without rule detectors. Node.js rejects unknown scan option keys with TypeError. Existing rule scanner APIs, finding order, and byte/codepoint/UTF-16 offset semantics are preserved. WASM remains rule-based. See the model session guide for configuration and setup.

​
Release validation

Publication requires both model backends to pass installed-package integration checks on Linux x86_64, Linux ARM64, macOS ARM64, and Windows for the release commit. macOS Intel checks still run and are advisory. Rust, Python, Node.js, and WASM publish from the same release commit.