scripts/verify_v4_backfill.py

"""
Simulate the WORST production case: LLM returns only 2 roles (drops BYJU's
and ML Edutech), no recruiter pitch, 3 bullets per role. Verify the v4
backfill + pitch enforcement + weaving brings it to 90%+.
"""
import os, sys, io, shutil
sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8')

src_pdf = r'C:\Users\Nxtwave\Desktop\resume\Saiteja_Tirunagari_Resume A 26 - Copy.pdf'
dst_pdf = 'data/resume/resume.pdf'
os.makedirs(os.path.dirname(dst_pdf), exist_ok=True)
if not os.path.exists(dst_pdf):
    shutil.copyfile(src_pdf, dst_pdf)

# Monkey-patch the LLM call to return a deliberately weak v4 response
from src.resume_customizer import ResumeCustomizer, _read_docx_text
from src.llm_client import LLMClient
from src.ats_scorer import score_resume


def make_weak_response(resume_dict, job_title, company):
    """Mimic worst-case LLM: 2 roles only, no pitch, 3 bullets each."""
    return {
        "name": resume_dict["name"],
        "contact": resume_dict["contact"],
        # NO recruiter pitch — generic 2-sentence summary
        "summary": (
            "Product Manager with 5+ years of experience driving product development "
            "in fast-paced startups. Built funnel optimization and AI-powered features "
            "that scaled to 140,000+ users with measurable business impact."
        ),
        "roles": [
            # Only 2 of the 4 original roles — LLM dropped older ones
            {
                "title": resume_dict["roles"][0]["title"],
                "company": resume_dict["roles"][0]["company"],
                "location": resume_dict["roles"][0]["location"],
                "dates": resume_dict["roles"][0]["dates"],
                "bullets": [
                    "Led NIAT Application Portal revamp scaling to 141,269 verified leads",
                    "Drove payment conversion lift of +35.87 percentage points",
                    "Built AI chatbot generating 6,776 leads",
                ],
            },
            {
                "title": resume_dict["roles"][1]["title"],
                "company": resume_dict["roles"][1]["company"],
                "location": resume_dict["roles"][1]["location"],
                "dates": resume_dict["roles"][1]["dates"],
                "bullets": [
                    "Managed 20 customer-success specialists covering 40,000 customers",
                    "Maintained refund rate below 5% and satisfaction above 95%",
                ],
            },
        ],
        "achievements": [
            "Scaled to 141,269 verified leads",
            "Lifted payment conversion +35.87 pp",
        ],
        "education": resume_dict["education"],
    }


# Patch LLMClient.tailor_resume_v4 to return our weak response
_original_v4 = LLMClient.tailor_resume_v4
def _mocked_v4(self, cfg, resume_dict, jd_text, job_title, company, assessment):
    return make_weak_response(resume_dict, job_title, company)
LLMClient.tailor_resume_v4 = _mocked_v4


from src.llm_client import LLMClient as LC
rc = ResumeCustomizer.__new__(ResumeCustomizer)
rc.llm = LC.__new__(LC)
rc.resume_text = ''
rc.output_dir = 'data/output/resumes/_v4_backfill_test'
os.makedirs(rc.output_dir, exist_ok=True)
rc.fast_model_cfg = {"model": "fake", "api_key": "fake", "base_url": "https://fake"}
rc._pending_summary_inject = []


print('━' * 78)
print(f'{"JD":<48} {"v4_only":>9} {"FinalATS":>9} {"Roles":>6} {"Words":>6}')
print('━' * 78)

jd_files = {
    'airtel_pm': ('Product Manager', 'Airtel'),
    'sumo_logic_pm': ('Product Manager', 'Sumo Logic'),
    'edgeverve_pm': ('Product Manager', 'EdgeVerve'),
    'aditya_birla_apm': ('Associate Product Manager', 'Aditya Birla Capital'),
}

for jd_file, (role_title, company) in jd_files.items():
    with open(f'tests/fixtures/jds/{jd_file}.txt', encoding='utf-8') as f:
        jd = f.read()
    job = {
        "title": role_title,
        "company": company,
        "description": jd,
        "_raw_assessment": {"ats_keywords": []},
    }
    filepath = os.path.join(rc.output_dir, f'{company.replace(" ", "_")}_{jd_file}.docx')

    # Call the actual v4 flow
    result = rc._generate_resume_v4(job, cfg=rc.fast_model_cfg, filepath=filepath)

    if not result:
        print(f'{(role_title + " @ " + company):<48} v4 FAILED')
        continue

    text = _read_docx_text(result)
    r = score_resume(text, jd)

    # Count roles + words in the actual generated DOCX
    from docx import Document
    doc = Document(result)
    role_lines = sum(
        1 for p in doc.paragraphs
        if p.text and any(co in p.text for co in ["NxtWave", "Think & Learn", "ML Edutech", "BYJU"])
    )
    print(f'{(role_title + " @ " + company):<48} {"weak":>9} {r["ats_score"]:>8}   {role_lines:>4}   {r["word_count"]:>5}')

print('━' * 78)
print('\nThis test simulates the WORST case: LLM returns 2 roles only, no')
print('pitch, 3 bullets each. v4 backfill + pitch enforcement + weaving')
print('should still produce 90%+ resumes with all 4 candidate roles preserved.')