Expand description
Built-in recognizer backends for the Gaze pseudonymization pipeline.
§Feature flags
| Feature | Default | What it enables |
|---|---|---|
phone-parser | yes | Parser-backed E.164 and national phone validation via phonenumber |
safety-net | no | NerSafetyNet observer pass |
safety-net-openai | no | OpenAI-filter safety net subprocess; also enables safety-net |
test-support | no | Fixture helpers for safety-net tests; not for production use |
With phone-parser disabled, parser-backed phone validators fail closed. There is
no silent regex-only fallback, so phone spans that require parser validation are not
detected.
Modules§
- safety_
net safety-net - validators
- Deterministic checksum validators re-exported for recognizer tests and adopters.
Structs§
- Anchored
Match Recognizer - Recognizes person names near literal locale cues.
- Dictionary
Recognizer - Lookup-based
Recognizerfor tenant-specific PII. - Label
Map - Labels file format.
- Locale
Aware Model Registry - Registry that resolves locale-aware model backends by locale fallback tier.
- Model
Hints - Execution hints supplied by the caller.
- Model
Input - Input passed to a locale-aware model.
- Model
Span - Span emitted by a locale-aware model.
- NerDetector
- NER detector backed by a pinned local model artifact set. Multiple
NerDetectorinstances with different backends may be stacked in the samePipeline; span-conflict resolution picks winners across detectors. - NerOptions
- NerRecognizer
- Regex
Detector - Regex-backed
Recognizerimplementation. - Verified
Artifacts - Verified artifact handles. Produced by
verify_artifacts, consumed byNerDetector::load. Split out so the load contract can be exercised by unit tests without initializing a backend runtime.
Enums§
- Anchored
Boundary - CuePosition
- Model
Error - Closed error set for locale-aware model routing and inference.
- Model
Stage - Pipeline stage requesting model inference.
- Name
Shape - NerBackend
Kind - Driver-style enum for NER backends. Backends are swappable under a common
NerBackendtrait; each owns its own model-specific state. MultipleNerDetectorinstances (e.g. a BERT token-classifier plus a GLiNER zero-shot model) can be stacked in the samePipeline— span-conflict resolution picks winners across all detectors. - NerLoad
Error - Normalizer
Kind - Recognizer
Error - Error type for recognizer construction and rulepack field validation.
- Region
phone-parser - Regions supported by national phone validators.
- Safety
Tier - Rulepack recognizer activation tier.
- Validator
Kind - Closed set of validator implementations used by validator-backed recognizers.
Traits§
- Locale
Aware Model - Locale-aware model backend for Pass-2 NER and Pass-3 safety-net inference.
Functions§
Type Aliases§
- Result
- Result alias for recognizer-local operations.