VacatureRadar/apps/jobs/services/employer_resolution.py

from __future__ import annotations

from dataclasses import dataclass, field
from difflib import SequenceMatcher
from urllib.parse import urlsplit

from apps.jobs.models import JobPosting
from apps.sources.models import Source

from .normalization import CanonicalJobDraft, normalize_token


MERGE_THRESHOLD = 0.96
TITLE_MIN_THRESHOLD = 0.92
CONFLICT_TITLE_THRESHOLD = 0.70
CONFLICT_EMPLOYER_THRESHOLD = 0.50
CONFLICT_LOCATION_THRESHOLD = 0.45


@dataclass(frozen=True)
class EmployerResolutionDecision:
    job: JobPosting | None
    reason: str
    confidence: float
    canonical_url: str | None = None
    conflict: bool = False
    evidence: list[str] = field(default_factory=list)


def _normalize_similarity(value: str) -> str:
    return normalize_token(value or "")


def _token_similarity(left: str, right: str) -> float:
    if not left or not right:
        return 0.0
    return SequenceMatcher(
        None, _normalize_similarity(left)[:12000], _normalize_similarity(right)[:12000]
    ).ratio()


def _weighted_similarity(
    title_score: float,
    location_score: float | None,
    employer_score: float | None,
    employer_domain_match: bool,
    canonical_host_match: bool,
) -> float:
    weights: list[tuple[float, float]] = [(title_score, 0.68), (employer_domain_match and 1.0 or 0.0, 0.12)]
    if location_score is not None:
        weights.append((location_score, 0.12))
    if employer_score is not None:
        weights.append((employer_score, 0.06))
    if canonical_host_match:
        weights.append((1.0, 0.05))
    total_weight = sum(weight for _, weight in weights)
    if total_weight == 0:
        return 0.0
    return sum(value * weight for value, weight in weights) / total_weight


def _domain_match(value: str, candidate: str) -> bool:
    if not value or not candidate:
        return False
    return _normalize_similarity(value).strip(".").lower() == _normalize_similarity(candidate).strip(".").lower()


def _host(value: str) -> str:
    return (urlsplit((value or "").lower()).hostname or "").strip(".")


def resolve_direct_employer_match(
    draft: CanonicalJobDraft, *, source: Source | None
) -> EmployerResolutionDecision:
    if source is None or source.source_type == Source.Type.EMPLOYER or not draft.normalized_title:
        return EmployerResolutionDecision(None, "no_direct_resolution", 0.0, canonical_url=None, conflict=False)

    candidates = JobPosting.objects.filter(
        status__in=[JobPosting.Status.ACTIVE, JobPosting.Status.NEW],
        direct_employer=True,
    ).select_related("employer").order_by("id")

    best: JobPosting | None = None
    best_score = 0.0
    best_conflict = False
    best_evidence: list[str] = []
    draft_host = _host(draft.canonical_url)

    for candidate in candidates:
        title_score = _token_similarity(draft.normalized_title, candidate.normalized_title)
        if title_score < TITLE_MIN_THRESHOLD:
            continue

        location_score: float | None = None
        if draft.location_text and candidate.raw_location:
            location_score = _token_similarity(draft.location_text, candidate.raw_location)

        employer_score: float | None = None
        if draft.employer_name and candidate.employer_name:
            employer_score = _token_similarity(draft.employer_name, candidate.employer_name)

        employer_domain_match = _domain_match(
            draft.employer_domain, candidate.employer.domain if candidate.employer else ""
        )
        canonical_host_match = _host(candidate.canonical_url) == draft_host

        score = _weighted_similarity(
            title_score=title_score,
            location_score=location_score,
            employer_score=employer_score,
            employer_domain_match=employer_domain_match,
            canonical_host_match=canonical_host_match,
        )

        has_conflict = False
        if draft.location_text and candidate.raw_location:
            if location_score is not None and location_score < CONFLICT_LOCATION_THRESHOLD:
                has_conflict = True
        if draft.employer_name and candidate.employer_name and (
            employer_score is not None and employer_score < CONFLICT_EMPLOYER_THRESHOLD
        ):
            has_conflict = True
        if draft.employer_name and not candidate.employer_name and title_score < CONFLICT_TITLE_THRESHOLD:
            has_conflict = True

        if score > best_score:
            best = candidate
            best_score = score
            best_conflict = has_conflict
            best_evidence = [
                f"title:{round(title_score, 3)}",
                f"location:{'none' if location_score is None else round(location_score, 3)}",
                f"employer:{'none' if employer_score is None else round(employer_score, 3)}",
                f"employer_domain:{int(employer_domain_match)}",
                f"canonical_host_match:{int(canonical_host_match)}",
            ]

    if best is None:
        return EmployerResolutionDecision(None, "no_direct_resolution", 0.0, conflict=False)

    if best_conflict:
        return EmployerResolutionDecision(
            None,
            "review_direct_conflict",
            best_score,
            canonical_url=best.canonical_url,
            conflict=True,
            evidence=best_evidence,
        )

    if best_score < MERGE_THRESHOLD:
        return EmployerResolutionDecision(
            None,
            "no_direct_resolution",
            best_score,
            canonical_url=None,
            conflict=False,
            evidence=best_evidence,
        )

    return EmployerResolutionDecision(
        best,
        "resolved_direct_match",
        best_score,
        canonical_url=best.canonical_url,
        conflict=False,
        evidence=best_evidence,
    )