Module ner 

Named Entity Recognition (NER) engine — CE-4 GLiNER zero-shot NER.

Two-layer extraction pipeline:

1. Rule-based pre-pass — regex extraction of dates, URLs, UUIDs, emails, IPs. Always on, zero latency, no model download required.
2. GLiNER ONNX engine — zero-shot NER via GLiNER-medium ONNX INT8 (52 MB). Opt-in per namespace, lazy-loaded on first use.

Extracted entities are stored as tags: entity:person:alice, entity:org:anthropic. Tag values are lowercased and whitespace-normalized for consistent deduplication.

Structs

ExtractedEntity

A single extracted entity.

GlinerEngine

GLiNER zero-shot NER engine backed by ONNX Runtime.

NerEngine

Unified NER engine combining rule-based and GLiNER extraction.

Functions

deduplicate_entities

Deduplicate entities by (entity_type, normalized_value), keeping the highest score.

normalize_label

Normalize an entity type label: lowercase, spaces → underscores.

rule_based_extract

Run the rule-based pre-pass — O(n) regex scan, zero model overhead.