wip: phase_z2 evidence 파이프라인 + matching 실험(phase2~26) + 프론트 trace 패널 진행분 스냅샷

- src: phase_z2 composition/mapper/pipeline/placement_planner/retry, ai_fallback(prompts/schema/validate), mdx_text_atoms 신규 - Front: PipelineTracePanel 신규, FramePanel/SlideCanvas/Home/designAgentApi 등 갱신 + 테스트 4종 추가 - templates/phase_z2: catalog(component_expansion_registry, node_slot_mapping 신규), frames, families, slide_base 갱신 - tests/matching: phase2~26 매칭 실험 스크립트·리포트·온톨로지 전체 (미커밋 진행분) - tests: b4_v4 evidence, task5~28.5 시리즈, regression(imp95 baseline) 등 신규 테스트 대량 추가 - docs/reference: MDX 구조 인벤토리, MDX→Frame 구조 계약 문서 - scripts: mdx 계약/parity/coverage/viewport 체크, gitea comment, run sync 유틸 - .gitignore: tmp*.json, chromedriver, .orchestrator, *.pkl, Front_test* 등 임시/스냅샷 제외 미완성 작업의 보존용 스냅샷 커밋 (2026-07-02) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-02 17:03:42 +09:00
parent 97b7833a1b
commit b836e79ee1
527 changed files with 673036 additions and 717 deletions
@@ -0,0 +1,145 @@
+"""Pipeline Step 17 — V4 (template-fit) 를 V3 Top-K 가 아니라 32 프레임 전체에 적용.
+
+기존 V4 r2 는 V3 Top-5 만 평가 → 5 개 중 선택.
+이 스크립트는 32 프레임 전체에 template-fit 적용 → confidence 기준 전체 랭킹.
+
+V1/V2/V3 처럼 "32 중 Top-3" 형태 산출.
+
+출력: v4_full32_result.yaml
+"""
+import datetime
+import sys
+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 template_fit import (
+    load_templates_v1, collect_anchor_vocab, compute_template_fit, route,
+)
+from embeddings import embed_texts, cosine
+from pipeline_01_extract_nodes import MDX_SECTIONS, MDX_DIR
+
+OUT_PATH = HERE / 'v4_full32_result.yaml'
+
+
+def extract_mdx_raw(sid):
+    cfg = MDX_SECTIONS[sid]
+    p = MDX_DIR / cfg['file']
+    lines = p.read_text(encoding='utf-8').split('\n')
+    start = None
+    for i, ln in enumerate(lines):
+        if ln.strip() == cfg['start'].strip():
+            start = i
+            break
+    end = len(lines)
+    if cfg.get('end_prefix'):
+        for i in range(start + 1, len(lines)):
+            if lines[i].strip().startswith(cfg['end_prefix']):
+                end = i
+                break
+    section = lines[start:end]
+    return section[0].lstrip('#').strip(), '\n'.join(section)
+
+
+def main():
+    v1 = yaml.safe_load((HERE / 'mdx_matching_result.yaml').read_text(encoding='utf-8'))
+    answer_map = v1['meta']['answer_map']
+    holdout = v1['meta']['holdout_sections']
+
+    templates = load_templates_v1()
+    anchor_vocab = collect_anchor_vocab(templates)
+    frames = load_32_frames()
+    idx_data, frame_to_short = load_frame_index()
+    fids = list(frames.keys())
+    frame_num_map = {fid: int(frame_to_short[fid]) for fid in fids}
+    frame_contents = [frames[fid].get('content', '') for fid in fids]
+
+    print('[V4 full-32] 32 frame content 임베딩 중...')
+    frame_vecs = embed_texts(frame_contents)
+
+    out_sections = {}
+    for sid in v1['mdx_sections']:
+        title, raw_text = extract_mdx_raw(sid)
+        mdx_analysis = detect_mdx_analysis(raw_text, title, anchor_vocab=anchor_vocab)
+        mdx_summary = mdx_analysis['summary']
+        mdx_vec = embed_texts([mdx_summary])[0]
+
+        judgments = []
+        for i, fid in enumerate(fids):
+            if fid not in templates:
+                continue
+            template = templates[fid]
+            content_emb = max(0.0, min(1.0, float(cosine(mdx_vec, frame_vecs[i]))))
+            fit = compute_template_fit(mdx_analysis, template, content_emb)
+            label = route(fit['confidence'], fit['axes'], fit['adaptation'], fit['not_suits'])
+            judgments.append({
+                'frame_id': fid,
+                'frame_number': frame_num_map[fid],
+                'template_id': template.get('template_id'),
+                'confidence': round(float(fit['confidence']), 4),
+                'base': round(float(fit['base']), 4),
+                'penalty': round(float(fit['total_penalty']), 4),
+                'label': label,
+                'content_embedding': round(content_emb, 4),
+                'axes': {
+                    'anchor': round(float(fit['axes']['anchor']['score']), 4),
+                    'cardinality': round(float(fit['axes']['cardinality']), 4),
+                    'relation': round(float(fit['axes']['relation']), 4),
+                    'slot': round(float(fit['axes']['slot']), 4),
+                    'content': round(float(fit['axes']['content']), 4),
+                },
+            })
+
+        # confidence 내림차순
+        judgments.sort(key=lambda x: -x['confidence'])
+        for new_rank, item in enumerate(judgments, start=1):
+            item['v4_full_rank'] = new_rank
+
+        ans = answer_map.get(sid)
+        out_sections[sid] = {
+            'mdx_title': title,
+            'answer_frame_number': ans,
+            'is_holdout': sid in holdout,
+            'judgments_full32': judgments,
+            'usable_count': sum(1 for j in judgments if j['label'] != 'reject'),
+            'reject_count': sum(1 for j in judgments if j['label'] == 'reject'),
+        }
+
+    out = {
+        'meta': {
+            'pipeline_step': '8.v4.full32',
+            'description': 'V4 template-fit 을 V3 Top-K 가 아니라 32 프레임 전체에 적용',
+            'note': '기존 v4_template_fit_r2_result.yaml 는 V3 Top-5 만 평가. 이 파일은 32 전체.',
+            'generated_at': datetime.datetime.now().isoformat(timespec='seconds'),
+            'answer_map': answer_map,
+            'holdout_sections': holdout,
+        },
+        'mdx_sections': out_sections,
+    }
+    OUT_PATH.write_text(
+        yaml.safe_dump(out, allow_unicode=True, sort_keys=False, width=1000),
+        encoding='utf-8',
+    )
+
+    print('=' * 70)
+    print(f'V4 full-32 평가 완료: {OUT_PATH}')
+    print('=' * 70)
+    print(f'\n각 섹션별 사용 가능 프레임 (label != reject) 개수 + Top-3:')
+    for sid, s in out_sections.items():
+        ans = s.get('answer_frame_number')
+        ans_str = f'(정답 {ans})' if ans else '(holdout)'
+        print(f'\n[{sid}] {ans_str}  사용가능 {s["usable_count"]}/{s["usable_count"]+s["reject_count"]}')
+        usable = [j for j in s['judgments_full32'] if j['label'] != 'reject']
+        for j in usable[:3]:
+            mark = ' 🎯' if ans and j['frame_number'] == ans else ''
+            print(f'  Frame {j["frame_number"]:>2}{mark}  conf {j["confidence"]:.3f}  {j["label"]}')
+        if not usable:
+            print('  (사용 가능 후보 없음 — 모두 reject)')
+
+
+if __name__ == '__main__':
+    main()