v0.5.0

What’s new

• In-memory text inputs: pass text records directly to Anonymizer without first creating an input file. (#322)
• Positive detection examples: guide entity detection with entity_label_examples. (#275)
• GLiNER2 span handling: preserve source text and offsets using overlapping native word-token windows, with bundled defaults of 128 tokens and 32-token overlap. Configured chunk_length and overlap values are now honored; omitted settings use the server defaults. (#282, #330)
• Rewrite recovery: retain generated rewrites when required replacements are unavailable, run normal evaluation and repair, and flag incomplete replacement application for human review. (#328)

Fixes and security updates

• Update dependencies and trust stores, including the eMudhra CA compromise response. (#320)
• Improve replacement configuration parsing and judge-response handling. (#303, #290)
• Correct private detection prompt configuration for DataDesigner. (#294)

Documentation

Updated notebooks, GLiNER2 self-hosting guidance, and quality and throughput tuning documentation. (#344, #338, #261)

Compatibility note: the bundled GLiNER2 server counts native word tokens, not characters or encoder subwords. Separately hosted detector services define their own chunk units and defaults.

Validation: 1,583 unit tests passed, repository checks and the strict documentation build passed, and RC installation from TestPyPI was verified.

Full changelog: v0.4.0...v0.5.0