"""Baseline : Jean-Baptiste/camembert-ner (generique PER/ORG/LOC/MISC).

Sert de point de comparaison dans la matrice (§5.3) : que vaut un NER FR
generique la ou Anonym-IA V2 est specialise PII ? Torch CPU via transformers,
pas d'export ONNX : c'est un candidat de matrice, pas un module produit.
"""

from functools import lru_cache

LABEL_MAP = {"PER": "PERSON", "ORG": "COMPANY", "LOC": "CITY", "MISC": "MISC"}


@lru_cache(maxsize=1)
def _pipe():
    from transformers import pipeline

    return pipeline("ner", model="Jean-Baptiste/camembert-ner",
                    aggregation_strategy="simple", device=-1)


def detect(text: str) -> list[dict]:
    spans = []
    for e in _pipe()(text):
        etype = LABEL_MAP.get(e["entity_group"], e["entity_group"])
        start, end = int(e["start"]), int(e["end"])
        # recalage frontieres de mots (meme raison que ner_onnx)
        while start > 0 and text[start].isalnum() and text[start - 1].isalnum():
            start -= 1
        while end < len(text) and text[end - 1].isalnum() and text[end].isalnum():
            end += 1
        spans.append({"start": start, "end": end, "type": etype,
                      "value": text[start:end]})
    return spans