apps/accounts/ner/lexical.py

"""Layer 0/floor: lexical catalog matching.

This wraps the existing pdfplumber-free matching logic from the MVP so the
dispatcher can treat it uniformly with the other layers. Unlike the four
ML layers, this one is always safe — it has no heavy deps and only touches
the DB's Skill catalog.

Kept separate from ``resume_parser`` so the dispatcher can invoke it as
a layer alongside the others rather than having a special-case code path.
"""
from __future__ import annotations

from . import NerLayer


class LexicalLayer(NerLayer):
    name = "lexical"

    def predict(self, text: str) -> dict[str, float]:
        if not text.strip():
            return {}
        # Reuse the existing implementation so behaviour stays identical.
        from apps.accounts.resume_parser import _candidate_matches
        from apps.skills.models import Skill

        skills = list(Skill.objects.all().only("id", "skill_name"))
        hits = _candidate_matches(text, skills)
        # _candidate_matches returns {skill_id: {confidence, proficiency, matched_span}}
        # — the dispatcher works with {name: confidence}, so remap.
        by_id = {s.id: s.skill_name for s in skills}
        out: dict[str, float] = {}
        for skill_id, data in hits.items():
            name = by_id.get(skill_id)
            if name:
                out[name] = float(data["confidence"])
        return out

    def available(self) -> bool:
        return True


layer = LexicalLayer()