apps/accounts/tests/test_resume_parser_integration.py

"""End-to-end NER chain integration test.

Gated on ``GAPGUIDE_ML_SMOKE=1`` — runs the REAL four-layer chain
(Nucha + JobBERT + SkillNER + SBERT + lexical) against a fixture PDF. CI
sets GAPGUIDE_PARSE_LAYERS=lexical (see conftest.py), so this opt-in test
is the only place we verify the ML stack actually works end-to-end.

Run manually:
    $env:GAPGUIDE_ML_SMOKE="1"
    pytest apps/accounts/tests/test_resume_parser_integration.py -q

First run downloads ~1 GB of HF models + requires `en_core_web_lg`
installed and pgvector extension in Postgres with SkillEmbedding populated.
"""
from __future__ import annotations

import os
from pathlib import Path

import pytest


pytestmark = [
    pytest.mark.django_db(transaction=True),
    pytest.mark.skipif(
        os.environ.get("GAPGUIDE_ML_SMOKE") != "1",
        reason="ML smoke disabled (set GAPGUIDE_ML_SMOKE=1 to enable)",
    ),
]


FIX = Path(__file__).resolve().parent.parent.parent.parent / 'tests' / 'fixtures' / 'resumes'


@pytest.fixture
def seeded_catalog():
    """Real catalog + SBERT embeddings. Requires:
      * pgvector extension on the test DB
      * en_core_web_lg installed
      * seed_initial_skills has been run (we do it here)
    """
    from django.core.management import call_command
    call_command('seed_initial_skills')

    # Build embeddings for the seeded skills.
    from apps.skills.models import Skill, SkillEmbedding
    from sentence_transformers import SentenceTransformer
    model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
    skills = list(Skill.objects.all())
    vectors = model.encode(
        [f"{s.skill_name} — {s.description or s.category}" for s in skills],
        normalize_embeddings=True,
    )
    for skill, vec in zip(skills, vectors):
        SkillEmbedding.objects.update_or_create(
            skill=skill,
            defaults={
                'embedding': vec.tolist(),
                'source_text': skill.skill_name,
            },
        )
    return skills


def test_full_chain_extracts_core_skills_from_strong_fixture(seeded_catalog):
    """resume_ds_strong.pdf must yield at least {Python, SQL} via the full chain.

    We deliberately don't assert exact counts / confidences — those depend
    on the specific HF model versions and would flake across upgrades.
    """
    # Force the full chain.
    os.environ.pop('GAPGUIDE_PARSE_LAYERS', None)
    from apps.accounts.resume_parser import parse_resume_envelope

    pdf = (FIX / 'resume_ds_strong.pdf').read_bytes()
    env = parse_resume_envelope(pdf)

    names = {s['skill_name'] for s in env['skills']}
    assert {'Python', 'SQL'}.issubset(names), (env['parser_version'], names)

    # At least one BERT/SBERT/SkillNER layer must have fired — proves the
    # chain isn't silently falling back to lexical-only.
    assert any(
        layer in env['parser_version']
        for layer in ('nucha', 'jobbert', 'skillner', 'sbert')
    ), env['parser_version']


def test_minimal_fixture_exercises_sbert_fallback(seeded_catalog):
    """resume_minimal.pdf uses paraphrased skills ("data visualisation",
    "version control"). Lexical layer alone returns nothing; the SBERT
    fallback should resolve at least one paraphrase to a catalog skill.
    """
    os.environ.pop('GAPGUIDE_PARSE_LAYERS', None)
    from apps.accounts.resume_parser import parse_resume_envelope

    pdf = (FIX / 'resume_minimal.pdf').read_bytes()
    env = parse_resume_envelope(pdf)

    # Either SBERT or SkillNER should salvage at least one catalog hit here —
    # otherwise the fallback chain isn't doing the job it was designed for.
    assert len(env['skills']) >= 1, (env['parser_version'], env['skills'])