from __future__ import annotations import argparse import asyncio import json import re import sys from pathlib import Path DESIGN_AGENT_ROOT = Path(r'D:\ad-hoc\kei\design_agent') if str(DESIGN_AGENT_ROOT) not in sys.path: sys.path.insert(0, str(DESIGN_AGENT_ROOT)) from src.block_reference import select_and_generate_references from src.config import settings from src.content_verifier import generate_with_retry import src.html_generator as html_generator from src.design_director import LAYOUT_PRESETS, select_preset from src.image_utils import embed_images, get_image_sizes from src.mdx_normalizer import normalize_mdx_content from src.pipeline_context import ( Analysis, BlockReference, ContainerInfo, DesignBudget, FontHierarchy, NormalizedContent, PageStructure, PipelineContext, Topic, create_context, ) from src.renderer import render_slide_from_html from src.slide_measurer import capture_slide_screenshot, measure_rendered_heights if not hasattr(html_generator, 'SIDEBAR_PROMPT') and hasattr(html_generator, '_LEGACY_SIDEBAR_PROMPT'): html_generator.SIDEBAR_PROMPT = html_generator._LEGACY_SIDEBAR_PROMPT if not hasattr(html_generator, 'FOOTER_PROMPT') and hasattr(html_generator, '_LEGACY_FOOTER_PROMPT'): html_generator.FOOTER_PROMPT = html_generator._LEGACY_FOOTER_PROMPT from src.space_allocator import ( ContainerSpec as LegacyContainerSpec, calculate_container_specs, calculate_design_budget, calculate_dynamic_ratio, calculate_font_hierarchy, ) def _load_json(path: Path) -> dict: return json.loads(path.read_text(encoding='utf-8-sig')) def _write_json(path: Path, data: dict) -> None: path.write_text(json.dumps(data, ensure_ascii=False, indent=2), encoding='utf-8') def _load_retry_plan(stage1b_path: Path) -> dict: retry_plan_path = stage1b_path.parent / 'retry-plan.json' if retry_plan_path.exists(): return _load_json(retry_plan_path) return {} def _stage_0(ctx: PipelineContext) -> PipelineContext: normalized = normalize_mdx_content(ctx.raw_content) ctx.normalized = NormalizedContent( clean_text=normalized['clean_text'], title=normalized['title'], images=normalized['images'], popups=normalized['popups'], tables=normalized['tables'], sections=normalized['sections'], ) ctx.save_snapshot('stage_0') return ctx def _stage_1a(ctx: PipelineContext, stage1a: dict) -> PipelineContext: analysis_raw = stage1a['analysis'] ctx.analysis = Analysis( core_message=analysis_raw['core_message'], title=analysis_raw['title'], total_pages=analysis_raw.get('total_pages', 1), ) ctx.page_structure = PageStructure(roles=stage1a['page_structure']) ctx.topics = [Topic(**raw) for raw in stage1a['topics']] ctx.save_snapshot('stage_1a') return ctx def _stage_1b(ctx: PipelineContext, stage1b: dict) -> PipelineContext: refined_map = {item['topic_id']: item for item in stage1b['concepts']} topics = [] for raw in ctx.topics: merged = raw.model_dump() if raw.id in refined_map: refined = dict(refined_map[raw.id]) refined.pop('source_data', None) merged.update(refined) topics.append(Topic(**merged)) ctx.topics = topics ctx.save_snapshot('stage_1b') return ctx def _stage_1_5a(ctx: PipelineContext) -> PipelineContext: image_sizes = get_image_sizes(ctx.raw_content, ctx.base_path) role_text_lengths = {} for role, info in ctx.page_structure.roles.items(): if isinstance(info, dict): role_text_lengths[role] = len(ctx.get_role_content(role)) font_hierarchy_dict = calculate_font_hierarchy(role_text_lengths) ctx.font_hierarchy = FontHierarchy( key_msg=font_hierarchy_dict.get('핵심', 14.0), core=font_hierarchy_dict.get('본심', 12.0), bg=font_hierarchy_dict.get('배경', 11.0), sidebar=font_hierarchy_dict.get('첨부', 10.0), ) ctx.container_ratio = calculate_dynamic_ratio(role_text_lengths, font_hierarchy_dict) analysis_dict = { 'topics': [t.model_dump() for t in ctx.topics], 'page_structure': ctx.page_structure.roles, } preset_name = select_preset(analysis_dict) ctx.preset_name = preset_name ctx.preset = LAYOUT_PRESETS.get(preset_name, {}) container_specs = calculate_container_specs( page_structure=ctx.page_structure.roles, topics=[t.model_dump() for t in ctx.topics], preset=ctx.preset, slide_width=settings.slide_width, slide_height=settings.slide_height, ) ctx.containers = { role: ContainerInfo( role=spec.role, zone=spec.zone, topic_ids=spec.topic_ids, weight=spec.weight, height_px=spec.height_px, width_px=spec.width_px, max_height_cost=spec.max_height_cost, block_constraints=spec.block_constraints, ) for role, spec in container_specs.items() } slide_images = [] for img_key, img_info in (image_sizes or {}).items(): img_path = Path(ctx.base_path) / img_key if ctx.base_path else Path(img_key) slide_images.append({ 'path': str(img_path), 'width': img_info.get('width', 0), 'height': img_info.get('height', 0), 'ratio': round(img_info.get('width', 1) / max(1, img_info.get('height', 1)), 2), 'topic_id': img_info.get('topic_id'), 'b64': '', }) ctx.slide_images = slide_images ctx.analysis = ctx.analysis.model_copy(update={'image_sizes': image_sizes or {}}) ctx.save_snapshot('stage_1_5a') return ctx def _stage_1_7(ctx: PipelineContext) -> PipelineContext: refs_raw = select_and_generate_references( topics=[t.model_dump() for t in ctx.topics], containers=ctx.containers, page_structure=ctx.page_structure.roles, ) normalized: dict[str, list[BlockReference]] = {} for role, ref in refs_raw.items(): ref_list = ref if isinstance(ref, list) else [ref] normalized[role] = [ BlockReference( block_id=item.get('block_id', ''), variant=item.get('variant', ''), visual_type=item.get('visual_type', ''), schema_info=item.get('schema_info', {}), design_reference_html=item.get('design_reference_html', ''), ) for item in ref_list if isinstance(item, dict) ] ctx.references = normalized ctx.save_snapshot('stage_1_7') return ctx def _stage_1_5b(ctx: PipelineContext) -> PipelineContext: updated = {} font_map = {'본심': 'core', '배경': 'bg', '첨부': 'sidebar', '결론': 'core'} for role, ci in ctx.containers.items(): refs = ctx.references.get(role, []) ref = refs[0] if refs else None schema_info = ref.schema_info if ref else {} font_size = getattr(ctx.font_hierarchy, font_map.get(role, 'core'), 12.0) budget = calculate_design_budget( container_height_px=ci.height_px, container_width_px=ci.width_px, block_schema=schema_info, font_size=font_size, ) updated[role] = ci.model_copy(update={ 'design_budget': DesignBudget( available_height_px=budget['available_height_px'], available_width_px=budget['available_width_px'], max_circle_diameter=budget['max_circle_diameter'], max_img_width=budget['max_img_width'], max_img_height=budget['max_img_height'], fits=budget['fits'], ) }) ctx.containers = updated ctx.save_snapshot('stage_1_5b') return ctx def _topic(ctx: PipelineContext, topic_id: int) -> Topic | None: return next((t for t in ctx.topics if t.id == topic_id), None) def compact_text(text: str, max_len: int) -> str: normalized = re.sub(r"\s+", " ", text).strip() if len(normalized) <= max_len: return normalized cut = normalized[:max_len].rsplit(" ", 1)[0].strip() return (cut or normalized[:max_len]).rstrip(" ,.;:") + "..." def preserve_80_percent(text: str, floor: int = 80, ceiling: int = 180) -> int: normalized = re.sub(r"\s+", " ", text).strip() if not normalized: return floor return max(floor, min(ceiling, int(len(normalized) * 0.8))) def _prefer_source_text(topic: Topic | None, fallback: str) -> str: if not topic: return fallback source = re.sub(r"\s+", " ", (topic.source_data or "")).strip() if source and len(source) >= max(80, len(fallback)): return source summary = re.sub(r"\s+", " ", (topic.summary or "")).strip() if source and len(source) >= 40: return source if summary: return summary return fallback def _trim_visible_copy(text: str, floor: int = 120, ceiling: int = 320) -> str: normalized = re.sub(r"\s+", " ", text).strip() if not normalized: return "" max_len = preserve_80_percent(normalized, floor=floor, ceiling=ceiling) return compact_text(normalized, max_len) def _extract_sentence(text: str, keyword: str, fallback: str) -> str: normalized = re.sub(r"\s+", " ", text).strip() if not normalized: return fallback parts = re.split(r"(?<=[.!?])\s+", normalized) for part in parts: if keyword in part: return part.strip() return fallback def _extract_multiple_sentences(text: str, keywords: list[str], fallback: str, limit: int = 2) -> str: normalized = re.sub(r"\s+", " ", text).strip() if not normalized: return fallback parts = [p.strip() for p in re.split(r"(?<=[.!?])\s+", normalized) if p.strip()] selected: list[str] = [] for keyword in keywords: for part in parts: if keyword in part and part not in selected: selected.append(part) break if len(selected) >= limit: break if selected: return " ".join(selected[:limit]) return fallback def _build_stage2_retry_html(ctx: PipelineContext, retry_plan: dict) -> dict: title = ctx.analysis.title problem_topic = _topic(ctx, 1) evidence_topic = _topic(ctx, 4) relation_topic = _topic(ctx, 3) comparison_topic = _topic(ctx, 5) conclusion_topic = _topic(ctx, 6) dx_topic = _topic(ctx, 2) problem_text = _trim_visible_copy(_prefer_source_text(problem_topic, '건설산업 디지털 전환 논의에서 DX와 BIM이 혼용되며 BIM 도입을 DX 완성으로 오인하는 문제가 발생하고 있다.'), floor=120, ceiling=260) relation_source = _prefer_source_text(relation_topic, 'DX와 GIS, BIM, Digital Twin의 관계를 시각적으로 드러낸다.') relation_text = _trim_visible_copy(relation_source, floor=110, ceiling=180) gis_line = _trim_visible_copy(_extract_sentence(relation_source, 'GIS', 'GIS는 공간 분석과 위치 기반 정보를 제공한다.'), floor=60, ceiling=140) bim_line = _trim_visible_copy(_extract_sentence(relation_source, 'BIM', 'BIM은 형상정보와 내용정보를 함께 다루는 핵심 인프라 기술이다.'), floor=60, ceiling=160) evidence_source = _prefer_source_text(evidence_topic, '정책 문서에서도 DX와 BIM이 혼용되며 이를 바로잡을 필요가 있다.') evidence_text = _trim_visible_copy(evidence_source, floor=90, ceiling=170) dx_source = _prefer_source_text(dx_topic, 'DX는 상위 개념이고 BIM은 이를 실행하는 핵심 기술이다.') dx_text = _trim_visible_copy(dx_source, floor=110, ceiling=220) compare_source = _prefer_source_text(comparison_topic, '범위·프로세스·성과품·확장성의 4개 비교축으로 DX와 BIM 차이를 짧고 직접적으로 보여준다.') compare_text = _trim_visible_copy(compare_source, floor=90, ceiling=120) conclusion_text = _trim_visible_copy(_prefer_source_text(conclusion_topic, '결론: BIM은 건설산업 DX를 수행하는 과정의 가장 기초가 되는 일부분이다.'), floor=70, ceiling=180) problem_title = problem_topic.title if problem_topic and problem_topic.title else 'DX와 BIM의 혼용 문제' dx_title = dx_topic.title if dx_topic and dx_topic.title else 'DX의 정의와 위치' relation_title = relation_topic.title if relation_topic and relation_topic.title else 'BIM과 핵심기술의 관계' comparison_title = comparison_topic.title if comparison_topic and comparison_topic.title else 'DX와 BIM 비교 핵심 포인트' evidence_title = evidence_topic.title if evidence_topic and evidence_topic.title else '정책 혼용 사례' body_html = f"""
{problem_title}
{problem_text}
{dx_title}
{dx_text}
[그림 1] DX와 핵심기술간 상호관계
DX
→
GIS
BIM
Digital Twin
{relation_title}
{relation_text}
GIS 역할
{gis_line}
BIM 역할
{bim_line}
{comparison_title}
{compare_text}
범위
DX는 BIM을 포함하는 상위 개념
프로세스
DX는 근본적 개선, BIM은 기존 2D 연장
성과품
DX는 공학 정보 연계, BIM은 3D 모델 중심
확장성
DX는 전 생애주기, BIM은 분야별 단절 위험
""".strip() sidebar_html = f"""
용어 정의
건설산업
다양한 기술을 통합해 시설물을 구현하는 종합 산업
BIM
3차원 모델 기반의 정보관리 도구이자 협업 인프라
출처: 국토교통부 BIM 기본지침
DX
디지털 기술 기반으로 업무방식과 가치구조를 전환하는 상위 개념
{evidence_title}
{evidence_text}
""".strip() footer_html = f"""
{conclusion_text}
""".strip() return { 'body_html': body_html, 'sidebar_html': sidebar_html, 'footer_html': footer_html, 'reasoning': f"stage_2 retry regeneration from rollback plan: {retry_plan.get('rollback_stage', 'stage_2')} with design-domain-guided slide composition", } async def _stage_2(ctx: PipelineContext, retry_plan: dict | None = None) -> PipelineContext: analysis_dict = { 'topics': [t.model_dump() for t in ctx.topics], 'page_structure': ctx.page_structure.roles, 'core_message': ctx.analysis.core_message, 'title': ctx.analysis.title, 'total_pages': ctx.analysis.total_pages, 'image_sizes': ctx.analysis.image_sizes, } container_specs_dict = { role: LegacyContainerSpec( role=ci.role, zone=ci.zone, topic_ids=ci.topic_ids, weight=ci.weight, height_px=ci.height_px, width_px=ci.width_px, max_height_cost=ci.max_height_cost, block_constraints=ci.block_constraints, ) for role, ci in ctx.containers.items() } analysis_dict['phase_t'] = { 'font_hierarchy': ctx.font_hierarchy.model_dump(), 'container_ratio': ctx.container_ratio, 'references': {role: [item.model_dump() for item in refs] for role, refs in ctx.references.items()}, 'design_budgets': { role: ci.design_budget.model_dump() if ci.design_budget else {} for role, ci in ctx.containers.items() }, } generated, verification = await generate_with_retry( content=ctx.raw_content, analysis=analysis_dict, container_specs=container_specs_dict, preset=ctx.preset, images=ctx.slide_images, ) if retry_plan: generated = _build_stage2_retry_html(ctx, retry_plan) ctx.generated_html = generated verification_path = ctx.get_run_dir() / 'stage_2_verification.json' _write_json(verification_path, { area: { 'passed': result.passed, 'score': result.score, 'errors': result.errors, } for area, result in verification.items() }) ctx.save_snapshot('stage_2') return ctx def _stage_3(ctx: PipelineContext) -> PipelineContext: analysis_dict = { 'topics': [t.model_dump() for t in ctx.topics], 'page_structure': ctx.page_structure.roles, 'core_message': ctx.analysis.core_message, 'title': ctx.analysis.title, } ctx.rendered_html = render_slide_from_html(ctx.generated_html, analysis_dict, ctx.preset) if ctx.base_path: ctx.rendered_html = embed_images(ctx.rendered_html, ctx.base_path) ctx.save_snapshot('stage_3') return ctx def _stage_4(ctx: PipelineContext) -> PipelineContext: ctx.measurement = measure_rendered_heights(ctx.rendered_html) ctx.screenshot_b64 = capture_slide_screenshot(ctx.rendered_html) or '' ctx.quality_score = 100 if not any( zone.get('overflowed') for zone in ctx.measurement.get('zones', {}).values() ) else 60 ctx.save_snapshot('stage_4') return ctx async def main() -> None: parser = argparse.ArgumentParser() parser.add_argument('--input', required=True) parser.add_argument('--stage1a', required=True) parser.add_argument('--stage1b', required=True) parser.add_argument('--base-path', default='') parser.add_argument('--output-dir', required=True) args = parser.parse_args() content = Path(args.input).read_text(encoding='utf-8') stage1a = _load_json(Path(args.stage1a)) stage1b_path = Path(args.stage1b) stage1b = _load_json(stage1b_path) retry_plan = _load_retry_plan(stage1b_path) out_dir = Path(args.output_dir) out_dir.mkdir(parents=True, exist_ok=True) ctx = create_context(content, args.base_path) ctx.run_dir = str(out_dir) ctx = _stage_0(ctx) ctx = _stage_1a(ctx, stage1a) ctx = _stage_1b(ctx, stage1b) ctx = _stage_1_5a(ctx) ctx = _stage_1_7(ctx) ctx = _stage_1_5b(ctx) ctx = await _stage_2(ctx, retry_plan=retry_plan or None) ctx = _stage_3(ctx) ctx = _stage_4(ctx) (out_dir / 'generated_html.json').write_text( json.dumps(ctx.generated_html, ensure_ascii=False, indent=2), encoding='utf-8', ) (out_dir / 'final.html').write_text(ctx.rendered_html, encoding='utf-8') (out_dir / 'measurement.json').write_text( json.dumps(ctx.measurement, ensure_ascii=False, indent=2), encoding='utf-8', ) (out_dir / 'context.json').write_text( ctx.model_dump_json(indent=2, exclude={'screenshot_b64', 'rendered_html'}), encoding='utf-8', ) (out_dir / 'final_context.json').write_text( ctx.model_dump_json(indent=2, exclude={'screenshot_b64', 'rendered_html'}), encoding='utf-8', ) if __name__ == '__main__': asyncio.run(main())