C.E.L_Slide_test2/tests/matching/pipeline_08_v2_semantic_rerank.py

"""Pipeline Step 8 — V2: V1 Top-K semantic rerank

파이프라인 위치:
  V1 (pipeline_06_2_mdx_matching.py) → **V2 (이 스크립트)** → V3 → V4

설계 원칙:
  - V1 결과(mdx_matching_result.yaml) 는 **고정 baseline**. 건드리지 않음.
  - V1 rank_by_matching_score Top-K 만 rerank 대상 — K=5 (기준 잠금).
  - 가중합 아님. **순수 rerank** — V1 점수는 저장만, 재정렬 키는 의미 유사도뿐.
  - AI 판단 없음. ko-sroberta 임베딩은 결정론적 (seed 고정 가정).

입력:
  - mdx_matching_result.yaml (V1)
  - MDX_SECTIONS 원문 (pipeline_01 의 MDX_DIR + section config)
  - Figma frame content (phase_common.load_32_frames() → analysis.md '내용')

처리:
  1. 각 MDX 섹션(TARGET 4 + Holdout 3)별로:
     - V1 Top-5 frame_id 추출
     - MDX summary = detect_mdx.build_summary (title + 첫 문단 + slot labels)
     - Top-5 frame.content ↔ summary 의 ko-sroberta cosine 계산
     - cosine 내림차순으로 재정렬
  2. lock_snapshot 기록 (V1 yaml + 이 스크립트 + 의존 모듈 sha256)

출력:
  - v2_semantic_rerank_result.yaml
"""
import hashlib
import sys
import datetime
from pathlib import Path
import yaml

HERE = Path(__file__).parent
sys.path.insert(0, str(HERE))

from phase_common import load_32_frames, load_frame_index
from detect_mdx import detect_mdx_analysis
from embeddings import embed_texts, cosine
from pipeline_01_extract_nodes import MDX_SECTIONS, MDX_DIR

TOP_K = 5
MODEL_ID = 'jhgan/ko-sroberta-multitask'

# ============================================================
# 상수 — lock_snapshot 대상
# ============================================================
LOCK_SNAPSHOT_FILES = [
    'pipeline_08_v2_semantic_rerank.py',
    'embeddings.py',
    'detect_mdx.py',
    'phase_common.py',
    'pipeline_01_extract_nodes.py',
    'synonyms.yaml',
    'mdx_matching_result.yaml',
]


def sha256_file(path: Path) -> str:
    return hashlib.sha256(path.read_bytes()).hexdigest()


def extract_mdx_raw_section(section_id: str) -> tuple[str, str]:
    """(title, body_text) 반환. title = 첫 줄 heading 마커 제거, body_text = 원문 그대로."""
    cfg = MDX_SECTIONS[section_id]
    p = MDX_DIR / cfg['file']
    lines = p.read_text(encoding='utf-8').split('\n')
    start_idx = None
    for i, ln in enumerate(lines):
        if ln.strip() == cfg['start'].strip():
            start_idx = i
            break
    if start_idx is None:
        raise ValueError(f"start_heading 찾지 못함: {cfg['start']!r} in {p}")
    end_idx = len(lines)
    if cfg.get('end_prefix'):
        for i in range(start_idx + 1, len(lines)):
            if lines[i].strip().startswith(cfg['end_prefix']):
                end_idx = i
                break
    section_lines = lines[start_idx:end_idx]
    raw = '\n'.join(section_lines)
    # title = heading 마커(#, ##, ###) 제거한 첫 줄
    title_line = section_lines[0].lstrip('#').strip()
    return title_line, raw


def main():
    # 1. V1 결과 로드
    v1_path = HERE / 'mdx_matching_result.yaml'
    v1 = yaml.safe_load(v1_path.read_text(encoding='utf-8'))

    # 2. Figma frame 로드 + content 임베딩 (한 번만)
    frames = load_32_frames()
    idx_data, frame_to_short = load_frame_index()
    fids = list(frames.keys())
    frame_contents = [frames[fid].get('content', '') for fid in fids]
    fid_to_idx = {fid: i for i, fid in enumerate(fids)}

    print(f"[V2] ko-sroberta 모델 로드 + 32 frame content 임베딩...")
    frame_vecs = embed_texts(frame_contents)

    # 3. 섹션별 처리
    out_sections = {}
    for sid, sec in v1['mdx_sections'].items():
        top = sec['rank_by_matching_score'][:TOP_K]
        top_fids = [r['frame_id'] for r in top]

        # MDX summary 생성
        title, raw_text = extract_mdx_raw_section(sid)
        analysis = detect_mdx_analysis(raw_text, title, anchor_vocab=None)
        summary = analysis['summary']

        # summary 임베딩
        mdx_vec = embed_texts([summary])[0]

        # Top-K 에 대해서만 cosine 계산
        rerank = []
        for v1_rank_idx, r in enumerate(top, start=1):
            fid = r['frame_id']
            frame_idx = fid_to_idx[fid]
            sem = cosine(mdx_vec, frame_vecs[frame_idx])
            rerank.append({
                'frame_id': fid,
                'frame_number': r['frame_number'],
                'v1_rank': v1_rank_idx,
                'v1_score': r['matching_score'],
                'semantic_score': round(float(sem), 4),
            })

        # semantic_score 내림차순 재정렬
        rerank.sort(key=lambda x: -x['semantic_score'])
        for new_rank, item in enumerate(rerank, start=1):
            item['v2_rank'] = new_rank

        out_sections[sid] = {
            'section_type': sec.get('section_type'),
            'answer_frame_number': sec.get('answer_frame_number'),
            'mdx_title': title,
            'mdx_summary': summary,
            'top_k': TOP_K,
            'v1_top_k': [
                {
                    'rank': i + 1,
                    'frame_id': r['frame_id'],
                    'frame_number': r['frame_number'],
                    'v1_score': r['matching_score'],
                }
                for i, r in enumerate(top)
            ],
            'v2_rerank': rerank,
        }

    # 4. lock_snapshot
    lock = {
        'timestamp': datetime.datetime.now().isoformat(timespec='seconds'),
        'model': MODEL_ID,
        'top_k': TOP_K,
        'files': {
            name: sha256_file(HERE / name) for name in LOCK_SNAPSHOT_FILES
        },
        'principle': [
            'V1 baseline 고정 — 재정렬 대상만 Top-K 로 한정',
            '가중합 금지 — 순수 semantic rerank',
            'Holdout 성적을 설계 근거로 사용하지 않음',
        ],
    }

    out = {
        'meta': {
            'pipeline_step': '8.v2',
            'description': 'V1 Top-K 후보를 ko-sroberta cosine 으로 재정렬 (캐스케이드 rerank 1단)',
            'model': MODEL_ID,
            'similarity': 'cosine',
            'top_k': TOP_K,
            'mdx_summary_spec': 'title + 첫 일반 문단(≤120자) + slot label Top-5 join (detect_mdx.build_summary)',
            'frame_source': 'analysis.md "내용" 섹션 (phase_common.load_32_frames)',
            'v1_source': 'mdx_matching_result.yaml rank_by_matching_score',
            'holdout_sections': v1['meta']['holdout_sections'],
            'answer_map': v1['meta']['answer_map'],
            'lock_snapshot': lock,
        },
        'mdx_sections': out_sections,
    }

    out_path = HERE / 'v2_semantic_rerank_result.yaml'
    out_path.write_text(
        yaml.safe_dump(out, allow_unicode=True, sort_keys=False, width=1000),
        encoding='utf-8',
    )

    # 콘솔 요약
    print()
    print("=" * 70)
    print(f"V2 재정렬 완료: {out_path}")
    print("=" * 70)
    answer_map = v1['meta']['answer_map']
    for sid, s in out_sections.items():
        answer_num = answer_map.get(sid)
        v1_top1 = s['v1_top_k'][0]['frame_number']
        v2_top1 = s['v2_rerank'][0]['frame_number']
        v2_top1_v1rank = s['v2_rerank'][0]['v1_rank']
        mark = ''
        if answer_num is not None:
            v1_ok = '✓' if v1_top1 == answer_num else '✗'
            v2_ok = '✓' if v2_top1 == answer_num else '✗'
            mark = f"  정답={answer_num}  V1{v1_ok}  V2{v2_ok}"
        else:
            mark = f"  (holdout)"
        print(f"  [{sid:8}] V1 top1=Frame {v1_top1:>2} → V2 top1=Frame {v2_top1:>2} (V1 rank {v2_top1_v1rank}){mark}")
    print()


if __name__ == '__main__':
    main()