Support per-field thresholds & regex validators in JSON schema (from_dict), plus per-task cls_threshold

[Andrew-Chen-Wang]

Motivation

Schema.from_dict / from_json are the way to build a schema from JSON (e.g. when GLiNER2 is served behind an HTTP API and clients send a SchemaInput payload). But the JSON format is missing three things the Python builder already supports, so they're unreachable from JSON:

• StructureBuilder.field(..., threshold=…) — per-field confidence threshold
• StructureBuilder.field(..., validators=[RegexValidator(...)]) — per-field regex span filters
• Schema.classification(..., cls_threshold=…) — per-task classification threshold

FieldInput only modeled name/dtype/choices/description and ClassificationInput only task/labels/multi_label, so from_dict silently dropped these capabilities.

Change

Wire the existing builder features through the JSON boundary:

• schema_model.py: new ValidatorInput model (pattern, mode, exclude, ignore_case) — a clean JSON mirror of RegexValidator that avoids exposing raw re flag ints. Add threshold + validators to FieldInput and cls_threshold to ClassificationInput, with range/regex validation.
• from_dict: pass threshold/validators into builder.field(...) (converting ValidatorInput → RegexValidator, ignore_case → re.IGNORECASE) and cls_threshold into classification(...).
• to_dict: emit the same keys (defaults omitted) so the dict round-trips through from_dict.
• Export ValidatorInput from the package; add torch-free tests.

Example

{
  "structures": {
    "contact": {
      "fields": [
        {"name": "name", "dtype": "str", "threshold": 0.7},
        {"name": "email", "dtype": "str",
         "validators": [{"pattern": "[^@]+@[^@]+", "mode": "partial", "ignore_case": false}]}
      ]
    }
  },
  "classifications": [
    {"task": "sentiment", "labels": ["positive", "negative"], "cls_threshold": 0.8}
  ]
}

Backward compatibility

All new fields are optional and default to the builders' existing defaults; to_dict omits them when unset, so existing schemas serialize byte-for-byte as before. from_dict(to_dict(schema)) is idempotent (covered by a test).

Tests

Added tests/test_schema_input_thresholds_validators.py (torch-free — no model download): from_dict wiring, to_dict round-trip/idempotency, defaults-omitted, and rejection of out-of-range thresholds / invalid regex.

🤖 Generated with Claude Code