C.E.L_Slide_test2/tests/matching/phase11.py

"""Phase 11 — 두 가지 신규 전략 비교
A. Asymmetric: MDX=제목만 × Figma=전체(전처리됨)
B. Ensemble: Track A(BM25 본문 × 전체) + Track B(Cross 제목 × 제목) 합산
"""
import sys
import json
from pathlib import Path

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

from common import load_figma_texts
from extract_units import extract_units
from methods import (
    method_tfidf, method_char_ngram, method_kiwi_bm25,
    method_distinctive_kiwi,
    method_sbert_chunk, method_e5_chunk, method_cross_encoder,
    method_hybrid_bm25_cross,
)
from phase10 import extract_titles_only_mdx, extract_titles_only_figma, TARGET_UNITS

ROOT = Path(r"d:\ad-hoc\kei\design_agent")
PREVIEW_DIR = ROOT / "data" / "figma_previews"


def short_id(fid, frame_to_short):
    return frame_to_short.get(str(fid), "?")


def normalize_scores(scored):
    """score list [(fid, score), ...] → 0-1 정규화"""
    if not scored:
        return {}
    max_s = max(s for _, s in scored)
    min_s = min(s for _, s in scored)
    if max_s == min_s:
        return {fid: 0.5 for fid, _ in scored}
    return {fid: (s - min_s) / (max_s - min_s) for fid, s in scored}


def method_ensemble(mdx_full, mdx_title, figma_full, figma_titles, alpha=0.5):
    """Track A: BM25(MDX 전체, Figma 전체) + Track B: Cross(MDX 제목, Figma 제목) 합산
    alpha: Track A 가중치 (1-alpha: Track B)"""
    track_a = method_kiwi_bm25(mdx_full, figma_full)
    track_b = method_cross_encoder(mdx_title, figma_titles)
    norm_a = normalize_scores(track_a)
    norm_b = normalize_scores(track_b)
    all_fids = set(norm_a.keys()) | set(norm_b.keys())
    combined = []
    for fid in all_fids:
        a = norm_a.get(fid, 0)
        b = norm_b.get(fid, 0)
        combined.append((fid, alpha * a + (1 - alpha) * b))
    return sorted(combined, key=lambda x: -x[1])


def main():
    units_full = extract_units()  # MDX 본문+제목 3x 가중
    figma_full = load_figma_texts()  # Figma 전체, 전처리+타이틀 가중
    figma_titles = extract_titles_only_figma()  # Figma 제목만

    # MDX 제목만 (Phase 10 포맷)
    mdx_titles = {}
    for uid, _, _, fname, mid, sub in TARGET_UNITS:
        mdx_titles[uid] = extract_titles_only_mdx(fname, mid, sub)

    with open(PREVIEW_DIR / "index.json", encoding="utf-8") as f:
        idx_data = json.load(f)
    frame_to_short = {info["frame_id"]: sid for sid, info in idx_data.items()}

    # 전략: (이름, fn(unit_id) → ranked list)
    def strategy_asymmetric(method_fn, name):
        def run(uid):
            return method_fn(mdx_titles[uid], figma_full)
        return (f"{name}: MDX제목 × Figma전체", run)

    def strategy_phase10(method_fn, name):
        # MDX 제목 × Figma 제목 (Phase 10)
        def run(uid):
            return method_fn(mdx_titles[uid], figma_titles)
        return (f"{name}: MDX제목 × Figma제목 (Phase 10)", run)

    def strategy_phase9(method_fn, name):
        def run(uid):
            return method_fn(units_full[uid], figma_full)
        return (f"{name}: MDX전체 × Figma전체 (Phase 9)", run)

    strategies = []
    # 1. Asymmetric (신규 A안)
    strategies.append(strategy_asymmetric(method_kiwi_bm25, "BM25"))
    strategies.append(strategy_asymmetric(method_cross_encoder, "Cross"))
    strategies.append(strategy_asymmetric(method_sbert_chunk, "SBERT-청크"))
    strategies.append(strategy_asymmetric(method_e5_chunk, "E5-청크"))
    strategies.append(strategy_asymmetric(lambda m, f: method_hybrid_bm25_cross(m, f, top_k=5), "BM25→Cross"))

    # 2. Ensemble (신규 B안) — alpha 3개로 샘플
    for a in [0.3, 0.5, 0.7]:
        name = f"Ensemble α={a} (A=BM25본문×전체 + B=Cross제목×제목)"
        def make_run(alpha):
            def run(uid):
                return method_ensemble(units_full[uid], mdx_titles[uid],
                                       figma_full, figma_titles, alpha=alpha)
            return run
        strategies.append((name, make_run(a)))

    # 비교용 베이스라인
    strategies.append(strategy_phase9(method_kiwi_bm25, "[기준] BM25"))
    strategies.append(strategy_phase9(method_cross_encoder, "[기준] Cross"))
    strategies.append(strategy_phase9(lambda m, f: method_hybrid_bm25_cross(m, f, top_k=5), "[기준] BM25→Cross"))

    print("=" * 70)
    print("Phase 11 — Asymmetric + Ensemble 실험")
    print("=" * 70)

    results_by_strategy = {}
    for sname, run_fn in strategies:
        top3_by = {}
        results = []
        for uid, _, correct_id, _, _, _ in TARGET_UNITS:
            try:
                ranked = run_fn(uid)
                top3 = [short_id(fid, frame_to_short) for fid, _ in ranked[:3]]
            except Exception as e:
                top3 = ["ERR", "-", "-"]
                print(f"ERR {sname} {uid}: {e}")
            top3_by[uid] = top3
            results.append(top3[0] == correct_id)
        score = sum(results)
        marks = ["✓" if r else "✗" for r in results]
        results_by_strategy[sname] = (results, top3_by)
        print(f"  {sname[:55]:55s}  {' '.join(marks)}  = {score}/4")

    # MD 리포트
    png_rel = "../../data/figma_previews/"
    lines = []
    lines.append("# Phase 11 — Asymmetric & Ensemble 실험")
    lines.append("")
    lines.append("**신규 A 안**: MDX 제목만 × Figma 전체(전처리됨) — 비대칭 매칭")
    lines.append("**신규 B 안**: BM25(MDX전체×Figma전체) + Cross(MDX제목×Figma제목) 점수 합산")
    lines.append("")

    lines.append("### 전략 × 4유닛 — 1순위 정답률")
    lines.append("")
    lines.append("| 전략 | MDX1 팝업(18) | MDX2 2.2(14) | MDX03-1(13) | MDX03-2(29) | 합계 |")
    lines.append("|------|---------------|--------------|-------------|-------------|------|")
    for sname, _ in strategies:
        results, top3_by = results_by_strategy[sname]
        marks = []
        for idx, (uid, _, correct_id, _, _, _) in enumerate(TARGET_UNITS):
            t3 = top3_by[uid]
            mark = "✓" if results[idx] else f"✗({t3[0]})"
            marks.append(mark)
        score = sum(results)
        lines.append(f"| {sname} | {marks[0]} | {marks[1]} | {marks[2]} | {marks[3]} | **{score}/4** |")
    lines.append("")

    # 유닛별 상세 (썸네일)
    lines.append("---")
    lines.append("")
    lines.append("## 유닛별 전략 Top-3 상세")
    lines.append("")
    for uid, display, correct_id, _, _, _ in TARGET_UNITS:
        lines.append(f"### {display} — 정답 Frame **{correct_id}**")
        lines.append("")
        lines.append("| 전략 | 1순위 | 2순위 | 3순위 |")
        lines.append("|------|-------|-------|-------|")
        for sname, _ in strategies:
            _, top3_by = results_by_strategy[sname]
            t3 = top3_by[uid]
            row = [sname]
            for sid in t3:
                info = idx_data.get(sid, {})
                png = info.get("png", "")
                title = info.get("title_text", "").strip().replace("\n", " ") or ""
                if len(title) > 15:
                    title = title[:15] + "…"
                mark = "⭐ " if sid == correct_id else ""
                cell = f"{mark}![{sid}]({png_rel}{png})<br>**{sid}**<br>{title}"
                row.append(cell)
            lines.append("| " + " | ".join(row) + " |")
        lines.append("")

    out_path = Path(__file__).parent / "MATRIX_PHASE11.md"
    out_path.write_text("\n".join(lines), encoding="utf-8")
    print(f"\n완료: {out_path}")


if __name__ == "__main__":
    main()