Add dimension QC, DZI generator, web viewer, and fix RAG query bug

- dimension_qc_checker.py: rules-based QC for dimension chains, overlaps, crowding
- generate_dzi.py: Deep Zoom Image tile pyramid generator for OpenSeadragon
- generate_web_viewer.py: OpenSeadragon viewer with SVG overlays and issue feedback buttons
- rag_query.py: fix LightRAG remove_think_tags crash on None response from LLM
- .gitignore: add *.pdf, *.db, backend/uploads/, backend/outputs/

.gitignore

@@ -12,6 +12,16 @@ __pycache__/
 # Secrets
 .env
 
+# Database
+*.db
+*.sqlite
+*.sqlite3
+
+# PDFs and uploads
+*.pdf
+backend/uploads/
+backend/outputs/
+
 # OS
 .DS_Store
 Thumbs.db

backend/README.md

@@ -0,0 +1,122 @@
+# Blueprint QC Backend
+
+FastAPI backend для сбора, хранения и разметки замечаний QC по чертежам.
+
+## Архитектура
+
+```
+┌─────────────┐     ┌──────────────┐     ┌─────────────┐
+│   React UI  │────▶│  FastAPI     │────▶│   SQLite    │
+│  (viewer)   │◀────│  Backend     │◀────│   Database  │
+└─────────────┘     └──────────────┘     └─────────────┘
+                           │
+                           ▼
+                    ┌──────────────┐
+                    │  OCR + QC    │
+                    │  Pipeline    │
+                    └──────────────┘
+```
+
+## Установка
+
+```bash
+cd backend
+pip install -r requirements.txt
+```
+
+## Запуск
+
+```bash
+python -m uvicorn app.main:app --reload --port 8000
+```
+
+API будет доступен на http://localhost:8000
+
+Документация (Swagger): http://localhost:8000/docs
+
+## API Endpoints
+
+### Проекты
+
+| Method | Endpoint | Описание |
+|--------|----------|----------|
+| POST | `/api/projects/upload` | Загрузка PDF, запуск обработки |
+| GET | `/api/projects` | Список проектов |
+| GET | `/api/projects/{id}` | Детали проекта |
+
+### Замечания (Issues)
+
+| Method | Endpoint | Описание |
+|--------|----------|----------|
+| GET | `/api/projects/{id}/issues` | Замечания проекта |
+| GET | `/api/issues/{id}` | Одно замечание |
+
+### Feedback (разметка)
+
+| Method | Endpoint | Описание |
+|--------|----------|----------|
+| POST | `/api/feedback` | Отметить TP/FP |
+| GET | `/api/feedback/stats` | Статистика |
+
+### Обучение
+
+| Method | Endpoint | Описание |
+|--------|----------|----------|
+| GET | `/api/training/data` | Данные для обучения |
+| POST | `/api/training/export` | Скачать JSON |
+| GET | `/api/stats` | Общая статистика |
+
+### Viewer
+
+| Method | Endpoint | Описание |
+|--------|----------|----------|
+| GET | `/viewer/{project_id}/{page}` | HTML viewer |
+| GET | `/viewer_tiles/{project_id}/{path}` | DZI тайлы |
+
+## Workflow
+
+1. **Загрузка**: `POST /api/projects/upload` с PDF
+2. **Обработка**: Backend запускает OCR → QC → DZI (фоново)
+3. **Просмотр**: `GET /viewer/{id}/{page}` — viewer с overlay
+4. **Разметка**: Пользователь кликает замечания → `POST /api/feedback` (is_true_positive: true/false)
+5. **Обучение**: `GET /api/training/data` — экспорт размеченных данных
+
+## Feedback Schema
+
+```json
+{
+  "issue_id": 123,
+  "is_true_positive": true,    // true = реальная проблема, false = ложное срабатывание
+  "comment": "Размер действительно плохо читается",
+  "action_taken": "fixed"      // fixed / ignored / not_sure
+}
+```
+
+## Training Data
+
+Каждый размеченный пример содержит:
+- `bbox`: координаты на PNG
+- `issue_type`: тип проблемы
+- `is_true_positive`: метка от пользователя
+- `image_path`: путь к PNG страницы
+- `dimension_text`: текст размера (если есть)
+- `confidence`: OCR confidence
+
+Накопив 100-200 размеченных примеров, можно:
+1. Fine-tune VLM (few-shot prompting)
+2. Обучить YOLO-детектор под ваши типы чертежей
+3. Дообучить правила QC (эвристики)
+
+## Модели БД
+
+### Project
+- id, name, pdf_filename, status, created_at, output_folder
+
+### Page
+- id, project_id, page_number, png_path, ocr_data, vlm_description
+
+### Issue
+- id, project_id, page_id, issue_type, severity, message, bbox, dimension_text, confidence
+
+### Feedback
+- id, issue_id, is_true_positive, comment, action_taken, created_at

backend/__init__.py

@@ -0,0 +1 @@
+# backend/__init__.py

backend/app/__init__.py

@@ -0,0 +1 @@
+# app/__init__.py

backend/app/crud.py

@@ -0,0 +1,192 @@
+# app/crud.py
+from sqlalchemy.orm import Session, joinedload
+from sqlalchemy import func
+from typing import List, Optional
+import app.models as models
+import app.schemas as schemas
+
+
+# ---------- Project ----------
+def create_project(db: Session, pdf_filename: str, name: Optional[str] = None) -> models.Project:
+    project = models.Project(
+        name=name or pdf_filename,
+        pdf_filename=pdf_filename,
+        status="uploaded"
+    )
+    db.add(project)
+    db.commit()
+    db.refresh(project)
+    return project
+
+def get_project(db: Session, project_id: int) -> Optional[models.Project]:
+    return db.query(models.Project).options(
+        joinedload(models.Project.pages),
+        joinedload(models.Project.issues).joinedload(models.Issue.feedback)
+    ).filter(models.Project.id == project_id).first()
+
+def get_projects(db: Session, skip: int = 0, limit: int = 100) -> List[models.Project]:
+    return db.query(models.Project).options(
+        joinedload(models.Project.pages),
+        joinedload(models.Project.issues).joinedload(models.Issue.feedback)
+    ).order_by(models.Project.created_at.desc()).offset(skip).limit(limit).all()
+
+def update_project_status(db: Session, project_id: int, status: str, error_message: str = None, output_folder: str = None):
+    project = get_project(db, project_id)
+    if project:
+        project.status = status
+        if error_message is not None:
+            project.error_message = error_message
+        # Очищаем ошибку при новом запуске или успехе
+        if status in ("processing", "completed") and error_message is None:
+            project.error_message = None
+        if output_folder:
+            project.output_folder = output_folder
+        if status == "completed":
+            from datetime import datetime
+            project.completed_at = datetime.utcnow()
+        db.commit()
+        db.refresh(project)
+    return project
+
+
+# ---------- Page ----------
+def create_page(db: Session, project_id: int, page_number: int, **kwargs) -> models.Page:
+    page = models.Page(project_id=project_id, page_number=page_number, **kwargs)
+    db.add(page)
+    db.commit()
+    db.refresh(page)
+    return page
+
+def get_page_by_number(db: Session, project_id: int, page_number: int) -> Optional[models.Page]:
+    return db.query(models.Page).filter(
+        models.Page.project_id == project_id,
+        models.Page.page_number == page_number
+    ).first()
+
+
+# ---------- Issue ----------
+def create_issue(db: Session, project_id: int, page_id: Optional[int], **kwargs) -> models.Issue:
+    issue = models.Issue(project_id=project_id, page_id=page_id, **kwargs)
+    db.add(issue)
+    db.commit()
+    db.refresh(issue)
+    return issue
+
+def get_issues(db: Session, project_id: Optional[int] = None, page_id: Optional[int] = None, 
+               severity: Optional[str] = None, issue_type: Optional[str] = None,
+               has_feedback: Optional[bool] = None, skip: int = 0, limit: int = 1000):
+    query = db.query(models.Issue)
+    if project_id:
+        query = query.filter(models.Issue.project_id == project_id)
+    if page_id:
+        query = query.filter(models.Issue.page_id == page_id)
+    if severity:
+        query = query.filter(models.Issue.severity == severity)
+    if issue_type:
+        query = query.filter(models.Issue.issue_type == issue_type)
+    if has_feedback is not None:
+        if has_feedback:
+            query = query.filter(models.Issue.feedback != None)
+        else:
+            query = query.filter(models.Issue.feedback == None)
+    return query.order_by(models.Issue.created_at.desc()).offset(skip).limit(limit).all()
+
+def get_issue(db: Session, issue_id: int) -> Optional[models.Issue]:
+    return db.query(models.Issue).filter(models.Issue.id == issue_id).first()
+
+
+# ---------- Feedback ----------
+def create_feedback(db: Session, feedback: schemas.FeedbackCreate) -> models.Feedback:
+    # Удалить старый feedback если есть
+    existing = db.query(models.Feedback).filter(models.Feedback.issue_id == feedback.issue_id).first()
+    if existing:
+        db.delete(existing)
+        db.commit()
+    
+    db_feedback = models.Feedback(**feedback.dict())
+    db.add(db_feedback)
+    db.commit()
+    db.refresh(db_feedback)
+    return db_feedback
+
+def get_feedback_stats(db: Session, project_id: Optional[int] = None):
+    query = db.query(models.Feedback)
+    if project_id:
+        query = query.join(models.Issue).filter(models.Issue.project_id == project_id)
+    
+    total = query.count()
+    true_positive = query.filter(models.Feedback.is_true_positive == True).count()
+    false_positive = query.filter(models.Feedback.is_true_positive == False).count()
+    unreviewed = query.filter(models.Feedback.is_true_positive == None).count()
+    
+    accuracy = true_positive / (true_positive + false_positive) if (true_positive + false_positive) > 0 else None
+    
+    return {
+        "total": total,
+        "true_positive": true_positive,
+        "false_positive": false_positive,
+        "unreviewed": unreviewed,
+        "accuracy_estimate": round(accuracy, 3) if accuracy else None
+    }
+
+
+# ---------- Stats ----------
+def get_stats(db: Session) -> schemas.StatsResponse:
+    total_projects = db.query(models.Project).count()
+    total_issues = db.query(models.Issue).count()
+    
+    # По типам
+    issue_types = db.query(models.Issue.issue_type, func.count(models.Issue.id)).group_by(models.Issue.issue_type).all()
+    issues_by_type = {t[0]: t[1] for t in issue_types}
+    
+    # Feedback
+    fb_stats = get_feedback_stats(db)
+    
+    return schemas.StatsResponse(
+        total_projects=total_projects,
+        total_issues=total_issues,
+        issues_by_type=issues_by_type,
+        feedback_stats={
+            "true_positive": fb_stats["true_positive"],
+            "false_positive": fb_stats["false_positive"],
+            "unreviewed": fb_stats["unreviewed"]
+        },
+        accuracy_estimate=fb_stats["accuracy_estimate"]
+    )
+
+
+# ---------- Training Data ----------
+def export_training_data(db: Session, project_id: Optional[int] = None, 
+                         only_labeled: bool = True) -> List[schemas.TrainingSample]:
+    """Экспорт данных для обучения ML-модели."""
+    query = db.query(models.Issue)
+    if project_id:
+        query = query.filter(models.Issue.project_id == project_id)
+    if only_labeled:
+        query = query.join(models.Feedback)
+    
+    issues = query.all()
+    samples = []
+    
+    for issue in issues:
+        sample = schemas.TrainingSample(
+            issue_id=issue.id,
+            issue_type=issue.issue_type,
+            severity=issue.severity,
+            message=issue.message,
+            bbox={
+                "x1": issue.bbox_x1,
+                "y1": issue.bbox_y1,
+                "x2": issue.bbox_x2,
+                "y2": issue.bbox_y2
+            },
+            dimension_text=issue.dimension_text,
+            confidence=issue.confidence,
+            page_number=issue.page.page_number if issue.page else None,
+            is_true_positive=issue.feedback.is_true_positive if issue.feedback else None,
+            image_path=issue.page.png_path if issue.page else None,
+            label=f"{'good' if issue.feedback and issue.feedback.is_true_positive else 'bad'}_{issue.issue_type.lower()}" if issue.feedback else None
+        )
+        samples.append(sample)
+    
+    return samples

backend/app/database.py

@@ -0,0 +1,23 @@
+# app/database.py
+import os
+from sqlalchemy import create_engine
+from sqlalchemy.ext.declarative import declarative_base
+from sqlalchemy.orm import sessionmaker
+
+SQLALCHEMY_DATABASE_URL = os.getenv("DATABASE_URL", "sqlite:///./blueprint_qc.db")
+
+engine = create_engine(
+    SQLALCHEMY_DATABASE_URL,
+    connect_args={"check_same_thread": False} if SQLALCHEMY_DATABASE_URL.startswith("sqlite") else {}
+)
+SessionLocal = sessionmaker(autocommit=False, autoflush=False, bind=engine)
+
+Base = declarative_base()
+
+
+def get_db():
+    db = SessionLocal()
+    try:
+        yield db
+    finally:
+        db.close()

backend/app/main.py

@@ -0,0 +1,401 @@
+# app/main.py
+"""
+FastAPI backend для Blueprint QC Service.
+
+API Endpoints:
+- POST /api/projects/upload       — загрузка PDF
+- GET  /api/projects              — список проектов
+- GET  /api/projects/{id}         — детали проекта
+- GET  /api/projects/{id}/issues  — замечания проекта
+- POST /api/feedback              — отметить замечание (TP/FP)
+- GET  /api/feedback/stats        — статистика feedback
+- GET  /api/training/data         — экспорт данных для обучения
+- GET  /api/stats                 — общая статистика
+- GET  /viewer/{project_id}/{page} — HTML viewer (серверный, без file://)
+"""
+
+import os
+import sys
+import re
+import json
+import shutil
+from pathlib import Path
+from typing import List, Optional
+from datetime import datetime
+
+from fastapi import FastAPI, File, UploadFile, Depends, HTTPException, Query
+from fastapi.responses import HTMLResponse, FileResponse, JSONResponse
+from fastapi.middleware.cors import CORSMiddleware
+from fastapi.staticfiles import StaticFiles
+from sqlalchemy.orm import Session
+
+# Добавить пути
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+sys.path.insert(0, str(Path(__file__).parent))
+
+from app.database import engine, get_db, Base
+from app import models, schemas, crud, processing
+
+# Создать таблицы
+Base.metadata.create_all(bind=engine)
+
+# Создать папку static если нет
+STATIC_DIR = Path(__file__).parent.parent / "static"
+STATIC_DIR.mkdir(exist_ok=True)
+
+app = FastAPI(
+    title="Blueprint QC API",
+    description="Сервис автоматической проверки чертежей и сбора данных для обучения",
+    version="0.1.0"
+)
+
+# CORS
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+# Пути
+# Абсолютные пути (backend запускается из разных директорий)
+BASE_DIR = Path(__file__).parent.parent  # backend/
+UPLOAD_DIR = BASE_DIR / "uploads"
+OUTPUT_DIR = BASE_DIR / "outputs"
+UPLOAD_DIR.mkdir(exist_ok=True)
+OUTPUT_DIR.mkdir(exist_ok=True)
+
+
+# ==================== PROJECTS ====================
+
+@app.post("/api/projects/upload", response_model=schemas.ProjectResponse)
+async def upload_pdf(
+    file: UploadFile = File(...),
+    name: Optional[str] = None,
+    db: Session = Depends(get_db)
+):
+    """Загрузка PDF файла. Запускает фоновую обработку."""
+    if not file.filename.endswith('.pdf'):
+        raise HTTPException(400, "Only PDF files allowed")
+    
+    # Сохранить файл
+    safe_name = Path(file.filename).name
+    pdf_path = UPLOAD_DIR / f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{safe_name}"
+    with open(pdf_path, "wb") as f:
+        shutil.copyfileobj(file.file, f)
+    
+    # Создать проект в БД (статус = uploaded, анализ запускается вручную)
+    project = crud.create_project(db, pdf_filename=safe_name, name=name or safe_name)
+    
+    # НЕ запускаем обработку автоматически — пользователь жмет "Анализировать"
+    # вручную, когда готов
+    
+    return project
+
+
+@app.get("/api/projects", response_model=List[schemas.ProjectDetail])
+def list_projects(skip: int = 0, limit: int = 100, db: Session = Depends(get_db)):
+    """Список всех проектов со страницами и замечаниями."""
+    return crud.get_projects(db, skip=skip, limit=limit)
+
+
+@app.get("/api/projects/{project_id}", response_model=schemas.ProjectDetail)
+def get_project(project_id: int, db: Session = Depends(get_db)):
+    """Детали проекта со страницами и замечаниями."""
+    project = crud.get_project(db, project_id)
+    if not project:
+        raise HTTPException(404, "Project not found")
+    return project
+
+
+@app.delete("/api/projects/{project_id}")
+def delete_project(project_id: int, db: Session = Depends(get_db)):
+    """Удалить проект и все связанные данные."""
+    project = crud.get_project(db, project_id)
+    if not project:
+        raise HTTPException(404, "Project not found")
+    
+    # Удалить файлы проекта
+    if project.output_folder:
+        import shutil
+        try:
+            shutil.rmtree(project.output_folder, ignore_errors=True)
+        except:
+            pass
+    
+    db.delete(project)
+    db.commit()
+    return {"status": "deleted", "project_id": project_id}
+
+
+@app.post("/api/projects/{project_id}/analyze")
+def analyze_project(project_id: int, use_tiling: bool = False, db: Session = Depends(get_db)):
+    """Запустить анализ (OCR + Layout + QC + DZI) для проекта."""
+    project = crud.get_project(db, project_id)
+    if not project:
+        raise HTTPException(404, "Project not found")
+    
+    if project.status == "processing":
+        return {"status": "already_processing", "project_id": project_id}
+    
+    if project.status == "completed":
+        pass
+    
+    pdf_files = list(UPLOAD_DIR.glob(f"*_{project.pdf_filename}"))
+    if not pdf_files:
+        pdf_files = list(UPLOAD_DIR.glob(project.pdf_filename))
+    
+    if not pdf_files:
+        raise HTTPException(404, f"PDF file not found in {UPLOAD_DIR}")
+    
+    pdf_path = pdf_files[0]
+    
+    crud.update_project_status(db, project_id, "processing")
+    
+    import threading
+    def run_in_background():
+        db_local = next(get_db())
+        try:
+            processing.run_pipeline(project.id, pdf_path, OUTPUT_DIR, db_local, use_tiling=use_tiling)
+        except Exception as e:
+            print(f"[ERROR] Pipeline failed for project {project.id}: {e}")
+            import traceback
+            traceback.print_exc()
+            crud.update_project_status(db_local, project.id, "error", error_message=str(e))
+        finally:
+            db_local.close()
+    
+    thread = threading.Thread(target=run_in_background)
+    thread.start()
+    
+    return {"status": "processing_started", "project_id": project_id, "ocr_engine": "tiling" if use_tiling else "standard"}
+
+
+# ==================== ISSUES ====================
+
+@app.get("/api/projects/{project_id}/issues")
+def get_issues(
+    project_id: int,
+    severity: Optional[str] = Query(None, description="error / warning / info"),
+    issue_type: Optional[str] = Query(None),
+    has_feedback: Optional[bool] = Query(None, description="Filter by feedback presence"),
+    db: Session = Depends(get_db)
+):
+    """Получить замечания проекта."""
+    issues = crud.get_issues(
+        db, project_id=project_id, severity=severity, 
+        issue_type=issue_type, has_feedback=has_feedback
+    )
+    
+    # Build response with page_number manually
+    result = []
+    for issue in issues:
+        item = {
+            "id": issue.id,
+            "project_id": issue.project_id,
+            "page_id": issue.page_id,
+            "page_number": issue.page.page_number if issue.page else None,
+            "issue_type": issue.issue_type,
+            "severity": issue.severity,
+            "message": issue.message,
+            "bbox_x1": issue.bbox_x1,
+            "bbox_y1": issue.bbox_y1,
+            "bbox_x2": issue.bbox_x2,
+            "bbox_y2": issue.bbox_y2,
+            "dimension_text": issue.dimension_text,
+            "confidence": issue.confidence,
+            "created_at": issue.created_at.isoformat() if issue.created_at else None,
+            "feedback": {
+                "id": issue.feedback.id,
+                "issue_id": issue.feedback.issue_id,
+                "is_true_positive": issue.feedback.is_true_positive,
+                "comment": issue.feedback.comment,
+                "action_taken": issue.feedback.action_taken,
+                "created_at": issue.feedback.created_at.isoformat() if issue.feedback.created_at else None
+            } if issue.feedback else None
+        }
+        result.append(item)
+    
+    return result
+
+
+@app.get("/api/issues/{issue_id}", response_model=schemas.IssueResponse)
+def get_issue(issue_id: int, db: Session = Depends(get_db)):
+    """Одно замечание."""
+    issue = crud.get_issue(db, issue_id)
+    if not issue:
+        raise HTTPException(404, "Issue not found")
+    return issue
+
+
+# ==================== FEEDBACK ====================
+
+@app.post("/api/feedback", response_model=schemas.FeedbackResponse)
+def submit_feedback(feedback: schemas.FeedbackCreate, db: Session = Depends(get_db)):
+    """
+    Отправить feedback по замечанию.
+    
+    is_true_positive:
+        - true  = реальная проблема (правильное срабатывание)
+        - false = ложное срабатывание (false positive)
+        - null  = не уверен
+    
+    action_taken: fixed / ignored / not_sure
+    """
+    issue = crud.get_issue(db, feedback.issue_id)
+    if not issue:
+        raise HTTPException(404, "Issue not found")
+    
+    return crud.create_feedback(db, feedback)
+
+
+@app.get("/api/feedback/stats")
+def feedback_stats(project_id: Optional[int] = None, db: Session = Depends(get_db)):
+    """Статистика feedback."""
+    return crud.get_feedback_stats(db, project_id=project_id)
+
+
+# ==================== VIEWER ====================
+
+@app.get("/viewer/{project_id}/{page_number}", response_class=HTMLResponse)
+def get_viewer(project_id: int, page_number: int, db: Session = Depends(get_db)):
+    """HTML viewer с overlay замечаний + feedback buttons."""
+    project = crud.get_project(db, project_id)
+    if not project or not project.output_folder:
+        raise HTTPException(404, "Project or output not found")
+    
+    # ВСЕГДА перегенерировать viewer для нужной страницы
+    # (иначе тайлы указывают на другую страницу)
+    result = processing.generate_viewer_html(db, project_id, page_number)
+    if not result:
+        raise HTTPException(404, "Viewer not available")
+    
+    viewer_path = Path(result)
+    content = viewer_path.read_text(encoding="utf-8")
+    
+    # Подменить пути тайлов — теперь они будут правильные
+    # (generate_web_viewer.py уже сгенерировал для page_{page_number:03d}_files/)
+    content = content.replace(
+        'Url: "./page_', 
+        f'Url: "/viewer_tiles/{project_id}/page_'
+    )
+    
+    # Внедрить project_id и page info для навигации
+    total_pages = len(project.pages) if project.pages else page_number
+    content = content.replace('const PROJECT_ID = null;', f'const PROJECT_ID = {project_id};')
+    content = content.replace('const TOTAL_PAGES = null;', f'const TOTAL_PAGES = {total_pages};')
+    content = content.replace('const CURRENT_PAGE = null;', f'const CURRENT_PAGE = {page_number};')
+    
+    # Обновить счётчик страниц в навбаре
+    content = content.replace(
+        f'<span id="totalPages">?</span>',
+        f'<span id="totalPages">{total_pages}</span>'
+    )
+    
+    # Получить issue IDs из БД для этой страницы
+    page = crud.get_page_by_number(db, project_id, page_number)
+    if page:
+        issues = crud.get_issues(db, project_id=project_id, page_id=page.id)
+        issue_db_ids = [str(i.id) for i in issues]
+        
+        # Внедрить DB IDs в HTML
+        content = content.replace('data-has-api="false"', 'data-has-api="true"')
+        
+        def inject_db_id(match):
+            idx = int(match.group(1)) - 1
+            if idx < len(issue_db_ids):
+                return match.group(0) + f' data-db-id="{issue_db_ids[idx]}"'
+            return match.group(0)
+        
+        content = re.sub(r'data-id="(\d+)"', inject_db_id, content)
+    
+    return HTMLResponse(content=content)
+
+
+@app.get("/viewer_tiles/{project_id}/{filename:path}")
+def get_tile(project_id: int, filename: str, db: Session = Depends(get_db)):
+    """Отдаёт DZI тайлы."""
+    project = crud.get_project(db, project_id)
+    if not project or not project.output_folder:
+        raise HTTPException(404)
+    
+    tile_path = Path(project.output_folder) / filename
+    if tile_path.exists() and str(tile_path).startswith(str(project.output_folder)):
+        return FileResponse(tile_path)
+    raise HTTPException(404)
+
+
+# ==================== TRAINING DATA ====================
+
+@app.get("/api/training/data", response_model=schemas.TrainingDataExport)
+def export_training_data(
+    project_id: Optional[int] = Query(None),
+    only_labeled: bool = Query(True),
+    format: str = Query("json", description="json / yolo / csv"),
+    db: Session = Depends(get_db)
+):
+    """
+    Экспорт данных для обучения ML-модели.
+    
+    only_labeled=true — только размеченные feedback'ом замечания.
+    """
+    samples = crud.export_training_data(db, project_id=project_id, only_labeled=only_labeled)
+    
+    if format == "json":
+        return schemas.TrainingDataExport(
+            total_samples=len(samples),
+            labeled_samples=len([s for s in samples if s.is_true_positive is not None]),
+            samples=samples,
+            export_format="json"
+        )
+    
+    # TODO: YOLO / CSV форматы
+    raise HTTPException(400, f"Format {format} not yet implemented")
+
+
+@app.post("/api/training/export")
+def download_training_export(
+    project_id: Optional[int] = None,
+    only_labeled: bool = True,
+    db: Session = Depends(get_db)
+):
+    """Скачать training data как JSON файл."""
+    samples = crud.export_training_data(db, project_id=project_id, only_labeled=only_labeled)
+    
+    export = {
+        "generated_at": datetime.utcnow().isoformat(),
+        "total_samples": len(samples),
+        "samples": [s.dict() for s in samples]
+    }
+    
+    return JSONResponse(content=export, media_type="application/json")
+
+
+# ==================== STATS ====================
+
+@app.get("/api/stats", response_model=schemas.StatsResponse)
+def get_stats(db: Session = Depends(get_db)):
+    """Общая статистика системы."""
+    return crud.get_stats(db)
+
+
+# ==================== HEALTH ====================
+
+@app.get("/api/health")
+def health_check():
+    return {"status": "ok", "version": "0.1.0"}
+
+
+# ==================== STATIC (для демо) ====================
+
+# Подключить статику
+try:
+    app.mount("/", StaticFiles(directory=str(STATIC_DIR), html=True), name="static")
+except RuntimeError:
+    pass  # Already mounted
+
+
+if __name__ == "__main__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=8000)

backend/app/models.py

@@ -0,0 +1,95 @@
+# app/models.py
+from sqlalchemy import Column, Integer, String, Float, DateTime, Text, ForeignKey, JSON, Boolean
+from sqlalchemy.orm import relationship
+from datetime import datetime
+from app.database import Base
+
+
+class Project(Base):
+    """Проект — один загруженный PDF."""
+    __tablename__ = "projects"
+    
+    id = Column(Integer, primary_key=True, index=True)
+    name = Column(String, index=True)
+    pdf_filename = Column(String, nullable=False)
+    status = Column(String, default="uploaded")  # uploaded / processing / completed / error
+    created_at = Column(DateTime, default=datetime.utcnow)
+    completed_at = Column(DateTime, nullable=True)
+    error_message = Column(Text, nullable=True)
+    
+    # Пути к файлам
+    output_folder = Column(String, nullable=True)
+    
+    pages = relationship("Page", back_populates="project", cascade="all, delete-orphan")
+    issues = relationship("Issue", back_populates="project", cascade="all, delete-orphan")
+
+
+class Page(Base):
+    """Страница PDF."""
+    __tablename__ = "pages"
+    
+    id = Column(Integer, primary_key=True, index=True)
+    project_id = Column(Integer, ForeignKey("projects.id"))
+    page_number = Column(Integer, nullable=False)
+    png_path = Column(String, nullable=True)
+    dzi_path = Column(String, nullable=True)
+    ocr_data = Column(JSON, nullable=True)  # full_ocr_results для этой страницы
+    vlm_description = Column(Text, nullable=True)
+    width = Column(Integer, nullable=True)
+    height = Column(Integer, nullable=True)
+    
+    project = relationship("Project", back_populates="pages")
+    issues = relationship("Issue", back_populates="page")
+
+
+class Issue(Base):
+    """Замечание QC — одна проблема на чертеже."""
+    __tablename__ = "issues"
+    
+    id = Column(Integer, primary_key=True, index=True)
+    project_id = Column(Integer, ForeignKey("projects.id"))
+    page_id = Column(Integer, ForeignKey("pages.id"))
+    
+    issue_type = Column(String, nullable=False)  # DIMENSION_OVERLAP, LOW_CONFIDENCE, etc.
+    severity = Column(String, nullable=False)    # error / warning / info
+    message = Column(Text, nullable=False)
+    
+    # Координаты bbox на PNG (в пикселях)
+    bbox_x1 = Column(Float, nullable=True)
+    bbox_y1 = Column(Float, nullable=True)
+    bbox_x2 = Column(Float, nullable=True)
+    bbox_y2 = Column(Float, nullable=True)
+    
+    # Дополнительные данные
+    dimension_text = Column(String, nullable=True)  # Текст размера (если применимо)
+    confidence = Column(Float, nullable=True)        # OCR confidence
+    extra_data = Column(JSON, nullable=True)         # Всё остальное
+    source = Column(String, nullable=True)           # "rules" или "vlm"
+    
+    created_at = Column(DateTime, default=datetime.utcnow)
+    
+    project = relationship("Project", back_populates="issues")
+    page = relationship("Page", back_populates="issues")
+    feedback = relationship("Feedback", back_populates="issue", uselist=False)
+
+
+class Feedback(Base):
+    """Feedback проектировщика — правда ли это замечание."""
+    __tablename__ = "feedback"
+    
+    id = Column(Integer, primary_key=True, index=True)
+    issue_id = Column(Integer, ForeignKey("issues.id"), unique=True)
+    
+    # True = реальная проблема, False = ложное срабатывание, None = не размечено
+    is_true_positive = Column(Boolean, nullable=True)
+    
+    # Почему (опционально)
+    comment = Column(Text, nullable=True)
+    
+    # Что сделал проектировщик
+    action_taken = Column(String, nullable=True)  # fixed / ignored / not_sure
+    
+    created_at = Column(DateTime, default=datetime.utcnow)
+    user_id = Column(String, nullable=True)  # Для многопользовательского режима
+    
+    issue = relationship("Issue", back_populates="feedback")

backend/app/processing.py

@@ -0,0 +1,217 @@
+# app/processing.py
+"""
+Интеграция с существующими скриптами OCR, QC, DZI.
+Запускает pipeline в фоне и сохраняет результаты в БД.
+"""
+
+import os
+import sys
+import json
+import re
+import subprocess
+import shutil
+from pathlib import Path
+from typing import Optional
+from sqlalchemy.orm import Session
+
+# Добавить корень проекта в path для импорта скриптов
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+import app.crud as crud
+import app.models as models
+
+
+def run_pipeline(project_id: int, pdf_path: Path, output_base: Path, db: Session, use_tiling: bool = False):
+    """Запускает полный pipeline: PDF → OCR → Layout → Elements → QC → DZI → БД."""
+    
+    project = crud.get_project(db, project_id)
+    if not project:
+        return
+    
+    crud.update_project_status(db, project_id, "processing")
+    
+    try:
+        # 1. Создать output папку
+        output_folder = output_base / f"project_{project_id}"
+        output_folder.mkdir(parents=True, exist_ok=True)
+        
+        script_dir = Path(__file__).parent.parent.parent  # backend/app -> backend -> opencode
+        
+        # 2. OCR + PNG (RapidOCR или Tiling OCR)
+        cmd = [sys.executable, str(script_dir / "process_any_pdf.py"), str(pdf_path), str(output_folder)]
+        if use_tiling:
+            cmd.append("--use-tiling")
+            print("[INFO] Tiling OCR enabled")
+        _run_command(cmd, cwd=str(script_dir))
+        
+        # 3. Layout Detection для каждой оригинальной страницы
+        # Только page_NNN.png, исключая визуализации (_dims, _layout и т.д.)
+        page_pngs = sorted([p for p in output_folder.glob("page_*.png")
+                           if not any(suffix in p.stem for suffix in ["_dims", "_layout", "_detected", "_preproc", "_ocr_compare"])])
+        for png in page_pngs:
+            ocr_json = output_folder / "full_ocr_results.json"
+            if ocr_json.exists():
+                try:
+                    _run_command([
+                        sys.executable, str(script_dir / "layout_detector.py"),
+                        str(png), str(ocr_json)
+                    ], cwd=str(script_dir))
+                    print(f"[INFO] Layout detection done for {png.name}")
+                except Exception as e:
+                    print(f"[WARN] Layout detection failed for {png.name}: {e}")
+        
+        # 4. Multi-Element Extraction (dimensions, positions, GOSTs, etc.)
+        for png in page_pngs:
+            ocr_json = output_folder / "full_ocr_results.json"
+            layout_json = output_folder / "layout.json"
+            if ocr_json.exists() and layout_json.exists():
+                try:
+                    _run_command([
+                        sys.executable, str(script_dir / "multi_element_extractor.py"),
+                        str(png), str(ocr_json), str(layout_json)
+                    ], cwd=str(script_dir))
+                    print(f"[INFO] Element extraction done for {png.name}")
+                except Exception as e:
+                    print(f"[WARN] Element extraction failed for {png.name}: {e}")
+        
+        # 5. QC (dimension_qc_checker.py) — правила
+        _run_command([
+            sys.executable, str(script_dir / "dimension_qc_checker.py"),
+            str(output_folder)
+        ], cwd=str(script_dir))
+        
+        # 6. DZI для каждой страницы
+        for png in page_pngs:
+            _run_command([
+                sys.executable, str(script_dir / "generate_dzi.py"),
+                str(png)
+            ], cwd=str(script_dir))
+        
+        # 7. Проверить результаты
+        ocr_path = output_folder / "full_ocr_results.json"
+        if not ocr_path.exists():
+            raise RuntimeError(f"OCR results not generated: {ocr_path}")
+        
+        # 8. Загрузить в БД
+        _import_results(db, project_id, output_folder)
+        
+        crud.update_project_status(db, project_id, "completed", output_folder=str(output_folder))
+        
+    except Exception as e:
+        crud.update_project_status(db, project_id, "error", error_message=str(e))
+        raise
+
+
+def _run_command(cmd: list, cwd: Optional[Path] = None):
+    """Запускает команду, проверяет exit code, выбрасывает исключение при ошибке."""
+    result = subprocess.run(
+        cmd,
+        capture_output=True,
+        text=True,
+        cwd=cwd
+    )
+    if result.returncode != 0:
+        stderr = result.stderr[:1000]
+        stdout = result.stdout[:500]
+        raise RuntimeError(f"Command failed ({result.returncode}): {' '.join(cmd)}\nSTDERR: {stderr}\nSTDOUT: {stdout}")
+    return result
+
+
+def _import_results(db: Session, project_id: int, output_folder: Path):
+    """Импорт OCR и QC результатов в БД. Очищает старые данные проекта перед импортом."""
+    
+    # Очистить старые страницы и замечания (cascade удалит issues и feedback)
+    db.query(models.Page).filter(models.Page.project_id == project_id).delete(synchronize_session=False)
+    db.query(models.Issue).filter(models.Issue.project_id == project_id).delete(synchronize_session=False)
+    db.commit()
+    
+    # Загрузить OCR
+    ocr_path = output_folder / "full_ocr_results.json"
+    if ocr_path.exists():
+        ocr = json.loads(ocr_path.read_text(encoding="utf-8"))
+        for page_data in ocr.get("pages", []):
+            page_num = page_data["page_number"]
+            png_path = output_folder / f"page_{page_num:03d}.png"
+            
+            page = crud.create_page(
+                db, project_id=project_id, page_number=page_num,
+                png_path=str(png_path) if png_path.exists() else None,
+                ocr_data=page_data
+            )
+    
+    # Загрузить VLM extraction descriptions в pages
+    vlm_path = output_folder / "vlm_extraction.json"
+    vlm_data = {}
+    if vlm_path.exists():
+        vlm_data = json.loads(vlm_path.read_text(encoding="utf-8"))
+    
+    # Обновить vlm_description для каждой страницы
+    for img_name, extraction in vlm_data.items():
+        page_num_match = re.search(r"page_(\d+)", img_name)
+        if page_num_match:
+            page_num = int(page_num_match.group(1))
+            page = crud.get_page_by_number(db, project_id, page_num)
+            if page:
+                desc = extraction.get("description", "")
+                page.vlm_description = desc[:2000] if desc else None  # ограничим длину
+                db.commit()
+    
+    # Загрузить QC issues (только rules — VLM issues удалены как ненадёжные)
+    qc_path = output_folder / "dimension_qc_report.json"
+    if qc_path.exists():
+        qc = json.loads(qc_path.read_text(encoding="utf-8"))
+        
+        for severity in ["errors", "warnings", "infos"]:
+            for item in qc.get(severity, []):
+                page_num = item.get("page")
+                if not page_num:
+                    continue
+                page = crud.get_page_by_number(db, project_id, page_num)
+                
+                # Извлечь bbox
+                bbox = item.get("bbox") or item.get("bbox1") or item.get("bbox_dim")
+                x1 = y1 = x2 = y2 = None
+                if bbox:
+                    if isinstance(bbox[0], list):
+                        xs = [p[0] for p in bbox]
+                        ys = [p[1] for p in bbox]
+                        x1, y1, x2, y2 = min(xs), min(ys), max(xs), max(ys)
+                    else:
+                        x1, y1, x2, y2 = bbox[0], bbox[1], bbox[2], bbox[3]
+                
+                crud.create_issue(
+                    db, project_id=project_id, page_id=page.id if page else None,
+                    issue_type=item.get("type", "UNKNOWN"),
+                    severity=item.get("severity", "warning"),
+                    message=item.get("message", ""),
+                    bbox_x1=x1, bbox_y1=y1, bbox_x2=x2, bbox_y2=y2,
+                    dimension_text=item.get("text"),
+                    confidence=item.get("confidence"),
+                    source="rules",
+                    extra_data={k: v for k, v in item.items() if k not in ["type", "severity", "message", "page", "text", "confidence", "bbox", "bbox1", "bbox2", "bbox_dim", "source"]}
+                )
+
+
+def generate_viewer_html(db: Session, project_id: int, page_number: int) -> Optional[str]:
+    """Генерирует HTML viewer для конкретной страницы."""
+    project = crud.get_project(db, project_id)
+    if not project or not project.output_folder:
+        return None
+    
+    output_folder = Path(project.output_folder)
+    
+    # Перегенерировать viewer для нужной страницы
+    # Запускаем из папки, где находится generate_web_viewer.py
+    script_dir = Path(__file__).parent.parent.parent  # backend/app -> backend -> opencode
+    result = _run_command([
+        sys.executable, str(script_dir / "generate_web_viewer.py"), str(output_folder), str(page_number)
+    ], cwd=str(script_dir))
+    
+    if result.returncode != 0:
+        print(f"[ERROR] generate_web_viewer.py failed: {result.stderr[:500]}")
+        return None
+    
+    viewer_path = output_folder / "web_viewer" / "index.html"
+    if viewer_path.exists():
+        return str(viewer_path)
+    return None

backend/app/schemas.py

@@ -0,0 +1,137 @@
+# app/schemas.py
+from pydantic import BaseModel
+from typing import List, Optional, Dict, Any
+from datetime import datetime
+
+
+# ---------- Project ----------
+class ProjectBase(BaseModel):
+    name: Optional[str] = None
+
+class ProjectCreate(ProjectBase):
+    pdf_filename: str
+
+class ProjectResponse(ProjectBase):
+    id: int
+    pdf_filename: str
+    status: str
+    created_at: datetime
+    completed_at: Optional[datetime] = None
+    error_message: Optional[str] = None
+    
+    class Config:
+        from_attributes = True
+
+class ProjectDetail(ProjectResponse):
+    pages: List["PageResponse"] = []
+    issues: List["IssueResponse"] = []
+    
+    class Config:
+        from_attributes = True
+
+
+# ---------- Page ----------
+class PageBase(BaseModel):
+    page_number: int
+    width: Optional[int] = None
+    height: Optional[int] = None
+
+class PageResponse(PageBase):
+    id: int
+    png_path: Optional[str] = None
+    vlm_description: Optional[str] = None
+    ocr_data: Optional[Dict[str, Any]] = None
+    issue_count: int = 0
+    
+    class Config:
+        from_attributes = True
+
+
+# ---------- Issue ----------
+class IssueBase(BaseModel):
+    issue_type: str
+    severity: str
+    message: str
+    page_number: Optional[int] = None
+    
+class IssueCreate(IssueBase):
+    bbox_x1: Optional[float] = None
+    bbox_y1: Optional[float] = None
+    bbox_x2: Optional[float] = None
+    bbox_y2: Optional[float] = None
+    dimension_text: Optional[str] = None
+    confidence: Optional[float] = None
+    extra_data: Optional[Dict[str, Any]] = None
+
+class IssueResponse(IssueBase):
+    id: int
+    project_id: int
+    bbox_x1: Optional[float] = None
+    bbox_y1: Optional[float] = None
+    bbox_x2: Optional[float] = None
+    bbox_y2: Optional[float] = None
+    dimension_text: Optional[str] = None
+    confidence: Optional[float] = None
+    source: Optional[str] = None
+    created_at: datetime
+    feedback: Optional["FeedbackResponse"] = None
+    page_id: Optional[int] = None
+    
+    class Config:
+        from_attributes = True
+
+
+# ---------- Feedback ----------
+class FeedbackCreate(BaseModel):
+    issue_id: int
+    is_true_positive: Optional[bool] = None
+    comment: Optional[str] = None
+    action_taken: Optional[str] = None  # fixed / ignored / not_sure
+
+class FeedbackResponse(BaseModel):
+    id: int
+    issue_id: int
+    is_true_positive: Optional[bool] = None
+    comment: Optional[str] = None
+    action_taken: Optional[str] = None
+    created_at: datetime
+    
+    class Config:
+        from_attributes = True
+
+
+# ---------- Stats ----------
+class StatsResponse(BaseModel):
+    total_projects: int
+    total_issues: int
+    issues_by_type: Dict[str, int]
+    feedback_stats: Dict[str, int]  # true_positive, false_positive, unreviewed
+    accuracy_estimate: Optional[float] = None  # Доля правильных срабатываний
+
+
+# ---------- Training Data ----------
+class TrainingSample(BaseModel):
+    """Один пример для обучения ML-модели."""
+    issue_id: int
+    issue_type: str
+    severity: str
+    message: str
+    bbox: Dict[str, float]  # x1, y1, x2, y2
+    dimension_text: Optional[str] = None
+    confidence: Optional[float] = None
+    page_number: int
+    is_true_positive: Optional[bool] = None
+    
+    # Для обучения детектора (YOLO)
+    image_path: Optional[str] = None  # Путь к PNG страницы
+    crop_path: Optional[str] = None  # Вырезанный фрагмент bbox
+    label: Optional[str] = None  # "good_dimension" / "bad_placement" / etc.
+
+class TrainingDataExport(BaseModel):
+    total_samples: int
+    labeled_samples: int  # Есть is_true_positive
+    samples: List[TrainingSample]
+    export_format: str  # json / yolo / csv
+    
+    class Config:
+        from_attributes = True

backend/backend.log

@@ -0,0 +1,223 @@
+INFO:     Started server process [49915]
+INFO:     Waiting for application startup.
+INFO:     Application startup complete.
+INFO:     Uvicorn running on http://0.0.0.0:8001 (Press CTRL+C to quit)
+INFO:     127.0.0.1:59330 - "GET /api/health HTTP/1.1" 200 OK
+[WARN] Command failed: /opt/homebrew/opt/python@3.11/bin/python3.11 generate_web_viewer.py /Users/kirillblinov/development/opencode/OCR/opencode/output_123 5
+       stderr: /opt/homebrew/Cellar/python@3.11/3.11.15_1/Frameworks/Python.framework/Versions/3.11/Resources/Python.app/Contents/MacOS/Python: can't open file '/Users/kirillblinov/development/opencode/OCR/generate_web_viewer.py': [Errno 2] No such file or directory
+
+INFO:     127.0.0.1:59332 - "GET /viewer/1/5 HTTP/1.1" 500 Internal Server Error
+ERROR:    Exception in ASGI application
+Traceback (most recent call last):
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/uvicorn/protocols/http/httptools_impl.py", line 421, in run_asgi
+    result = await app(  # type: ignore[func-returns-value]
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/uvicorn/middleware/proxy_headers.py", line 56, in __call__
+    return await self.app(scope, receive, send)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/applications.py", line 1159, in __call__
+    await super().__call__(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/applications.py", line 90, in __call__
+    await self.middleware_stack(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/errors.py", line 186, in __call__
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/errors.py", line 164, in __call__
+    await self.app(scope, receive, _send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/cors.py", line 88, in __call__
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/exceptions.py", line 63, in __call__
+    await wrap_app_handling_exceptions(self.app, conn)(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app
+    await app(scope, receive, sender)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/middleware/asyncexitstack.py", line 18, in __call__
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 660, in __call__
+    await self.middleware_stack(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 680, in app
+    await route.handle(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 276, in handle
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 134, in app
+    await wrap_app_handling_exceptions(app, request)(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app
+    await app(scope, receive, sender)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 120, in app
+    response = await f(request)
+               ^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 674, in app
+    raw_response = await run_endpoint_function(
+                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 330, in run_endpoint_function
+    return await run_in_threadpool(dependant.call, **values)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/concurrency.py", line 32, in run_in_threadpool
+    return await anyio.to_thread.run_sync(func)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/to_thread.py", line 63, in run_sync
+    return await get_async_backend().run_sync_in_worker_thread(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/_backends/_asyncio.py", line 2518, in run_sync_in_worker_thread
+    return await future
+           ^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/_backends/_asyncio.py", line 1002, in run
+    result = context.run(func, *args)
+             ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/development/opencode/OCR/opencode/backend/app/main.py", line 285, in get_viewer
+    content = re.sub(r'data-id="(\d+)"', inject_db_id, content)
+              ^^
+NameError: name 're' is not defined
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/14/33_5.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/14/33_4.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/13/17_2.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/13/17_1.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/13/17_3.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/13/17_0.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /viewer_tiles/1/page_005_files/13/17_4.png HTTP/1.1" 200 OK
+[WARN] Command failed: /opt/homebrew/opt/python@3.11/bin/python3.11 generate_web_viewer.py /Users/kirillblinov/development/opencode/OCR/opencode/output_123 5
+       stderr: /opt/homebrew/Cellar/python@3.11/3.11.15_1/Frameworks/Python.framework/Versions/3.11/Resources/Python.app/Contents/MacOS/Python: can't open file '/Users/kirillblinov/development/opencode/OCR/generate_web_viewer.py': [Errno 2] No such file or directory
+
+INFO:     127.0.0.1:59372 - "GET /viewer/1/5 HTTP/1.1" 500 Internal Server Error
+ERROR:    Exception in ASGI application
+Traceback (most recent call last):
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/uvicorn/protocols/http/httptools_impl.py", line 421, in run_asgi
+    result = await app(  # type: ignore[func-returns-value]
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/uvicorn/middleware/proxy_headers.py", line 56, in __call__
+    return await self.app(scope, receive, send)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/applications.py", line 1159, in __call__
+    await super().__call__(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/applications.py", line 90, in __call__
+    await self.middleware_stack(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/errors.py", line 186, in __call__
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/errors.py", line 164, in __call__
+    await self.app(scope, receive, _send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/cors.py", line 88, in __call__
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/exceptions.py", line 63, in __call__
+    await wrap_app_handling_exceptions(self.app, conn)(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app
+    await app(scope, receive, sender)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/middleware/asyncexitstack.py", line 18, in __call__
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 660, in __call__
+    await self.middleware_stack(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 680, in app
+    await route.handle(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 276, in handle
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 134, in app
+    await wrap_app_handling_exceptions(app, request)(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app
+    await app(scope, receive, sender)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 120, in app
+    response = await f(request)
+               ^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 674, in app
+    raw_response = await run_endpoint_function(
+                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 330, in run_endpoint_function
+    return await run_in_threadpool(dependant.call, **values)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/concurrency.py", line 32, in run_in_threadpool
+    return await anyio.to_thread.run_sync(func)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/to_thread.py", line 63, in run_sync
+    return await get_async_backend().run_sync_in_worker_thread(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/_backends/_asyncio.py", line 2518, in run_sync_in_worker_thread
+    return await future
+           ^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/_backends/_asyncio.py", line 1002, in run
+    result = context.run(func, *args)
+             ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/development/opencode/OCR/opencode/backend/app/main.py", line 285, in get_viewer
+    content = re.sub(r'data-id="(\d+)"', inject_db_id, content)
+              ^^
+NameError: name 're' is not defined
+INFO:     127.0.0.1:59355 - "GET /api/projects HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59406 - "GET /api/stats HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59355 - "GET /api/projects HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59408 - "GET /api/stats HTTP/1.1" 200 OK
+[WARN] Command failed: /opt/homebrew/opt/python@3.11/bin/python3.11 generate_web_viewer.py /Users/kirillblinov/development/opencode/OCR/opencode/output_123 5
+       stderr: /opt/homebrew/Cellar/python@3.11/3.11.15_1/Frameworks/Python.framework/Versions/3.11/Resources/Python.app/Contents/MacOS/Python: can't open file '/Users/kirillblinov/development/opencode/OCR/generate_web_viewer.py': [Errno 2] No such file or directory
+
+INFO:     127.0.0.1:59425 - "GET /viewer/1/5 HTTP/1.1" 500 Internal Server Error
+ERROR:    Exception in ASGI application
+Traceback (most recent call last):
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/uvicorn/protocols/http/httptools_impl.py", line 421, in run_asgi
+    result = await app(  # type: ignore[func-returns-value]
+             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/uvicorn/middleware/proxy_headers.py", line 56, in __call__
+    return await self.app(scope, receive, send)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/applications.py", line 1159, in __call__
+    await super().__call__(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/applications.py", line 90, in __call__
+    await self.middleware_stack(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/errors.py", line 186, in __call__
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/errors.py", line 164, in __call__
+    await self.app(scope, receive, _send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/cors.py", line 88, in __call__
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/middleware/exceptions.py", line 63, in __call__
+    await wrap_app_handling_exceptions(self.app, conn)(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app
+    await app(scope, receive, sender)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/middleware/asyncexitstack.py", line 18, in __call__
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 660, in __call__
+    await self.middleware_stack(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 680, in app
+    await route.handle(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/routing.py", line 276, in handle
+    await self.app(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 134, in app
+    await wrap_app_handling_exceptions(app, request)(scope, receive, send)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 53, in wrapped_app
+    raise exc
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/_exception_handler.py", line 42, in wrapped_app
+    await app(scope, receive, sender)
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 120, in app
+    response = await f(request)
+               ^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 674, in app
+    raw_response = await run_endpoint_function(
+                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/fastapi/routing.py", line 330, in run_endpoint_function
+    return await run_in_threadpool(dependant.call, **values)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/starlette/concurrency.py", line 32, in run_in_threadpool
+    return await anyio.to_thread.run_sync(func)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/to_thread.py", line 63, in run_sync
+    return await get_async_backend().run_sync_in_worker_thread(
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/_backends/_asyncio.py", line 2518, in run_sync_in_worker_thread
+    return await future
+           ^^^^^^^^^^^^
+  File "/Users/kirillblinov/Library/Python/3.11/lib/python/site-packages/anyio/_backends/_asyncio.py", line 1002, in run
+    result = context.run(func, *args)
+             ^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/Users/kirillblinov/development/opencode/OCR/opencode/backend/app/main.py", line 285, in get_viewer
+    content = re.sub(r'data-id="(\d+)"', inject_db_id, content)
+              ^^
+NameError: name 're' is not defined
+INFO:     127.0.0.1:59553 - "GET /viewer_tiles/1/page_005_files/14/18_1.png HTTP/1.1" 200 OK
+INFO:     127.0.0.1:59553 - "GET /viewer_tiles/1/page_005_files/14/18_0.png HTTP/1.1" 200 OK
+INFO:     Shutting down
+INFO:     Waiting for application shutdown.
+INFO:     Application shutdown complete.
+INFO:     Finished server process [49915]

backend/import_existing.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Импорт существующего output_123 в backend БД.
+
+Использование:
+    python import_existing.py
+"""
+
+import sys
+import json
+from pathlib import Path
+
+# Добавить backend в path
+sys.path.insert(0, str(Path(__file__).parent.parent / "backend"))
+
+from app.database import SessionLocal, Base, engine
+from app import models, crud, schemas
+
+
+def import_output_123():
+    """Импорт существующего проекта output_123 в БД."""
+    
+    output_folder = Path("/Users/kirillblinov/development/opencode/OCR/opencode/output_123")
+    if not output_folder.exists():
+        print(f"[ERR] Не найден: {output_folder}")
+        return
+    
+    db = SessionLocal()
+    
+    try:
+        # Создать проект
+        project = crud.create_project(
+            db, 
+            pdf_filename="123.pdf",
+            name="Test Project 123"
+        )
+        print(f"[INFO] Создан проект: ID={project.id}")
+        
+        # Обновить статус
+        crud.update_project_status(db, project.id, "completed", output_folder=str(output_folder))
+        
+        # Загрузить OCR
+        ocr_path = output_folder / "full_ocr_results.json"
+        if ocr_path.exists():
+            ocr = json.loads(ocr_path.read_text(encoding="utf-8"))
+            
+            # Создать страницы
+            for page_data in ocr.get("pages", []):
+                page_num = page_data["page_number"]
+                png_path = output_folder / f"page_{page_num:03d}.png"
+                
+                from PIL import Image
+                width = height = None
+                if png_path.exists():
+                    with Image.open(png_path) as img:
+                        width, height = img.size
+                
+                page = crud.create_page(
+                    db,
+                    project_id=project.id,
+                    page_number=page_num,
+                    png_path=str(png_path) if png_path.exists() else None,
+                    ocr_data=page_data,
+                    width=width,
+                    height=height
+                )
+                print(f"  [OK] Страница {page_num}: {width}x{height}")
+        
+        # Загрузить QC issues
+        qc_path = output_folder / "dimension_qc_report.json"
+        if qc_path.exists():
+            qc = json.loads(qc_path.read_text(encoding="utf-8"))
+            
+            total_imported = 0
+            for severity in ["errors", "warnings", "infos"]:
+                for item in qc.get(severity, []):
+                    page_num = item["page"]
+                    page = crud.get_page_by_number(db, project.id, page_num)
+                    
+                    # Извлечь bbox
+                    bbox = item.get("bbox") or item.get("bbox1") or item.get("bbox_dim")
+                    x1 = y1 = x2 = y2 = None
+                    if bbox:
+                        if isinstance(bbox[0], list):
+                            xs = [p[0] for p in bbox]
+                            ys = [p[1] for p in bbox]
+                            x1, y1, x2, y2 = min(xs), min(ys), max(xs), max(ys)
+                        else:
+                            x1, y1, x2, y2 = bbox[0], bbox[1], bbox[2], bbox[3]
+                    
+                    crud.create_issue(
+                        db,
+                        project_id=project.id,
+                        page_id=page.id if page else None,
+                        issue_type=item["type"],
+                        severity=item["severity"],
+                        message=item["message"],
+                        bbox_x1=x1, bbox_y1=y1, bbox_x2=x2, bbox_y2=y2,
+                        dimension_text=item.get("text"),
+                        confidence=item.get("confidence"),
+                        extra_data={k: v for k, v in item.items() 
+                                   if k not in ["type", "severity", "message", "page", "text", "confidence", "bbox", "bbox1", "bbox2", "bbox_dim"]}
+                    )
+                    total_imported += 1
+            
+            print(f"[OK] Импортировано замечаний: {total_imported}")
+        
+        # Загрузить VLM descriptions
+        vlm_path = output_folder / "vlm_descriptions.json"
+        if vlm_path.exists():
+            vlm = json.loads(vlm_path.read_text(encoding="utf-8"))
+            for page in project.pages:
+                for item in vlm.get("descriptions", []):
+                    if item.get("page") == page.page_number:
+                        page.vlm_description = item.get("description")
+                        db.commit()
+                        break
+            print(f"[OK] VLM описания загружены")
+        
+        print(f"\n[INFO] Проект #{project.id} готов!")
+        print(f"      Dashboard: http://localhost:8001/")
+        print(f"      Viewer: http://localhost:8001/viewer/{project.id}/5")
+        print(f"      Review: http://localhost:8001/review.html?project={project.id}")
+        
+        return project.id
+        
+    except Exception as e:
+        print(f"[ERR] {e}")
+        import traceback
+        traceback.print_exc()
+    finally:
+        db.close()
+
+
+if __name__ == "__main__":
+    import_output_123()

backend/requirements.txt

@@ -0,0 +1,6 @@
+fastapi>=0.104.0
+uvicorn[standard]>=0.24.0
+sqlalchemy>=2.0.0
+pydantic>=2.5.0
+python-multipart>=0.0.6
+pillow>=10.0.0

backend/static/index.html

@@ -0,0 +1,329 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Blueprint QC Dashboard</title>
+    <style>
+        * { margin: 0; padding: 0; box-sizing: border-box; }
+        body {
+            font-family: 'Segoe UI', system-ui, sans-serif;
+            background: #0d0d1a;
+            color: #eee;
+            min-height: 100vh;
+        }
+        .container { max-width: 1200px; margin: 0 auto; padding: 40px 20px; }
+        h1 { color: #e94560; margin-bottom: 30px; }
+        .stats {
+            display: grid;
+            grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+            gap: 15px;
+            margin-bottom: 40px;
+        }
+        .stat-card {
+            background: #16213e;
+            border: 1px solid #0f3460;
+            padding: 20px;
+            border-radius: 8px;
+        }
+        .stat-value { font-size: 32px; font-weight: bold; color: #e94560; }
+        .stat-label { font-size: 13px; opacity: 0.7; margin-top: 8px; }
+        .upload-zone {
+            border: 2px dashed #0f3460;
+            border-radius: 12px;
+            padding: 60px 40px;
+            text-align: center;
+            margin-bottom: 40px;
+            transition: all 0.3s;
+        }
+        .upload-zone:hover, .upload-zone.dragover {
+            border-color: #e94560;
+            background: rgba(233, 69, 96, 0.05);
+        }
+        .spinner {
+            display: inline-block;
+            width: 16px;
+            height: 16px;
+            border: 2px solid rgba(255,170,0,0.3);
+            border-top-color: #ffaa00;
+            border-radius: 50%;
+            animation: spin 1s linear infinite;
+            vertical-align: middle;
+            margin-right: 6px;
+        }
+        @keyframes spin { to { transform: rotate(360deg); } }
+        .processing-row { animation: pulse 2s infinite; }
+        @keyframes pulse { 0%, 100% { opacity: 1; } 50% { opacity: 0.7; } }
+        .delete-btn {
+            background: transparent;
+            border: none;
+            color: #ff4444;
+            cursor: pointer;
+            font-size: 16px;
+            margin-left: 10px;
+            opacity: 0.5;
+        }
+        .delete-btn:hover { opacity: 1; }
+        .upload-zone input { display: none; }
+        .upload-btn {
+            background: #e94560;
+            color: white;
+            border: none;
+            padding: 12px 32px;
+            border-radius: 6px;
+            cursor: pointer;
+            font-size: 16px;
+            margin-top: 15px;
+        }
+        .projects-table {
+            width: 100%;
+            border-collapse: collapse;
+        }
+        .projects-table th {
+            text-align: left;
+            padding: 15px;
+            border-bottom: 2px solid #0f3460;
+            font-size: 13px;
+            text-transform: uppercase;
+            opacity: 0.7;
+        }
+        .projects-table td {
+            padding: 15px;
+            border-bottom: 1px solid #1a1a2e;
+        }
+        .projects-table tr:hover { background: rgba(255,255,255,0.03); }
+        .status {
+            display: inline-block;
+            padding: 4px 12px;
+            border-radius: 20px;
+            font-size: 12px;
+            font-weight: 600;
+        }
+        .status-completed { background: rgba(0,200,100,0.2); color: #00c864; }
+        .status-processing { background: rgba(255,170,0,0.2); color: #ffaa00; }
+        .status-error { background: rgba(255,0,0,0.2); color: #ff4444; }
+        .status-uploaded { background: rgba(100,100,255,0.2); color: #8888ff; }
+        .view-link { color: #44aaff; text-decoration: none; }
+        .view-link:hover { text-decoration: underline; }
+        #progress-bar {
+            display: none;
+            width: 100%;
+            height: 4px;
+            background: #1a1a2e;
+            border-radius: 2px;
+            margin-top: 15px;
+            overflow: hidden;
+        }
+        #progress-fill {
+            width: 0%;
+            height: 100%;
+            background: #e94560;
+            transition: width 0.3s;
+        }
+        .error-msg { color: #ff4444; margin-top: 10px; font-size: 14px; }
+    </style>
+</head>
+<body>
+    <div class="container">
+        <h1>Blueprint QC Dashboard</h1>
+        
+        <div class="stats" id="stats">
+            <div class="stat-card">
+                <div class="stat-value" id="stat-projects">-</div>
+                <div class="stat-label">Проектов</div>
+            </div>
+            <div class="stat-card">
+                <div class="stat-value" id="stat-issues">-</div>
+                <div class="stat-label">Замечаний</div>
+            </div>
+            <div class="stat-card">
+                <div class="stat-value" id="stat-feedback">-</div>
+                <div class="stat-label">Размечено</div>
+            </div>
+            <div class="stat-card">
+                <div class="stat-value" id="stat-accuracy">-</div>
+                <div class="stat-label">Точность</div>
+            </div>
+        </div>
+        
+        <div class="upload-zone" id="uploadZone">
+            <h3>Перетащите PDF сюда или кликните для выбора</h3>
+            <p style="opacity:0.6; margin-top:10px;">Максимальный размер: 100 МБ</p>
+            <input type="file" id="fileInput" accept=".pdf">
+            <button class="upload-btn" onclick="document.getElementById('fileInput').click()">Выбрать файл</button>
+            <div id="progress-bar"><div id="progress-fill"></div></div>
+            <div id="upload-error" class="error-msg"></div>
+        </div>
+        
+        <h2 style="margin-bottom:20px;">Проекты</h2>
+        <table class="projects-table" id="projectsTable">
+            <thead>
+                <tr>
+                    <th>ID</th>
+                    <th>Название</th>
+                    <th>Файл</th>
+                    <th>Статус</th>
+                    <th>Создан</th>
+                    <th>Замечаний</th>
+                    <th>Действия</th>
+                </tr>
+            </thead>
+            <tbody></tbody>
+        </table>
+    </div>
+
+    <script>
+        const API = '/api';
+        
+        // Stats
+        async function loadStats() {
+            try {
+                const res = await fetch(`${API}/stats`);
+                const data = await res.json();
+                document.getElementById('stat-projects').textContent = data.total_projects;
+                document.getElementById('stat-issues').textContent = data.total_issues;
+                document.getElementById('stat-feedback').textContent = data.feedback_stats.true_positive + data.feedback_stats.false_positive;
+                document.getElementById('stat-accuracy').textContent = data.accuracy_estimate ? (data.accuracy_estimate * 100).toFixed(1) + '%' : '-';
+            } catch (e) { console.error('Stats error:', e); }
+        }
+        
+        // Track previous statuses for detecting completion
+        let prevStatuses = {};
+        
+        // Projects
+        async function loadProjects() {
+            try {
+                const res = await fetch(`${API}/projects`);
+                const projects = await res.json();
+                const tbody = document.querySelector('#projectsTable tbody');
+                
+                // Check for newly completed projects
+                projects.forEach(p => {
+                    if (prevStatuses[p.id] === 'processing' && p.status === 'completed') {
+                        showToast(`Project "${p.name || p.pdf_filename}" обработан!`, 'success');
+                    }
+                    prevStatuses[p.id] = p.status;
+                });
+                
+                tbody.innerHTML = projects.map(p => {
+                    const isProcessing = p.status === 'processing' || p.status === 'uploaded';
+                    const statusHtml = isProcessing 
+                        ? `<span class="status status-${p.status}"><span class="spinner"></span>${p.status}</span>`
+                        : `<span class="status status-${p.status}">${p.status}</span>`;
+                    let actions;
+                    if (p.status === 'completed') {
+                        actions = `<a href="/viewer/${p.id}/1" class="view-link" target="_blank">Viewer</a> | <a href="/review.html?project=${p.id}" class="view-link">Review</a> <button class="delete-btn" onclick="deleteProject(${p.id})" title="Удалить">🗑</button>`;
+                    } else if (p.status === 'processing') {
+                        actions = '<span style="opacity:0.5"><span class="spinner"></span> Обработка...</span>';
+                    } else {
+                        actions = `
+                            <label style="font-size:11px; margin-right:6px; cursor:pointer;">
+                                <input type="checkbox" id="tiling-${p.id}" style="vertical-align:middle;"> tiling
+                            </label>
+                            <button class="upload-btn" style="padding:4px 12px; font-size:12px;" onclick="analyzeProject(${p.id})">🔬 Анализировать</button>
+                            <button class="delete-btn" onclick="deleteProject(${p.id})" title="Удалить">🗑</button>
+                        `;
+                    }
+                    return `
+                    <tr class="${isProcessing ? 'processing-row' : ''}">
+                        <td>${p.id}</td>
+                        <td>${p.name || '-'}</td>
+                        <td>${p.pdf_filename}</td>
+                        <td>${statusHtml}</td>
+                        <td>${new Date(p.created_at).toLocaleString()}</td>
+                        <td>${p.issues ? p.issues.length : (isProcessing ? '<span class="spinner"></span>' : '0')}</td>
+                        <td>${actions}</td>
+                    </tr>
+                `}).join('');
+            } catch (e) { console.error('Projects error:', e); }
+        }
+        
+        async function deleteProject(id) {
+            if (!confirm('Удалить проект?')) return;
+            try {
+                const res = await fetch(`${API}/projects/${id}`, { method: 'DELETE' });
+                if (res.ok) {
+                    loadProjects();
+                    loadStats();
+                    showToast('Проект удалён', 'success');
+                }
+            } catch(e) { showToast('Ошибка удаления', 'error'); }
+        }
+        
+        // Upload
+        const uploadZone = document.getElementById('uploadZone');
+        const fileInput = document.getElementById('fileInput');
+        const progressBar = document.getElementById('progress-bar');
+        const progressFill = document.getElementById('progress-fill');
+        const uploadError = document.getElementById('upload-error');
+        
+        uploadZone.addEventListener('dragover', (e) => { e.preventDefault(); uploadZone.classList.add('dragover'); });
+        uploadZone.addEventListener('dragleave', () => uploadZone.classList.remove('dragover'));
+        uploadZone.addEventListener('drop', (e) => {
+            e.preventDefault();
+            uploadZone.classList.remove('dragover');
+            if (e.dataTransfer.files.length) handleFile(e.dataTransfer.files[0]);
+        });
+        
+        fileInput.addEventListener('change', (e) => { if (e.target.files.length) handleFile(e.target.files[0]); });
+        
+        async function analyzeProject(id) {
+            try {
+                const useTiling = document.getElementById(`tiling-${id}`)?.checked || false;
+                const url = useTiling ? `${API}/projects/${id}/analyze?use_tiling=true` : `${API}/projects/${id}/analyze`;
+                const res = await fetch(url, { method: 'POST' });
+                if (res.ok) {
+                    const data = await res.json();
+                    showToast(`Анализ запущен (${data.ocr_engine})!`, 'success');
+                    loadProjects();
+                } else {
+                    const err = await res.text();
+                    showToast('Ошибка запуска: ' + err, 'error');
+                }
+            } catch(e) { showToast('Ошибка сети', 'error'); }
+        }
+        
+        async function handleFile(file) {
+            if (!file.name.endsWith('.pdf')) {
+                uploadError.textContent = 'Только PDF файлы';
+                return;
+            }
+            uploadError.textContent = '';
+            progressBar.style.display = 'block';
+            progressFill.style.width = '30%';
+            
+            const form = new FormData();
+            form.append('file', file);
+            form.append('name', file.name);
+            
+            try {
+                const res = await fetch(`${API}/projects/upload`, { method: 'POST', body: form });
+                progressFill.style.width = '100%';
+                if (res.ok) {
+                    setTimeout(() => { progressBar.style.display = 'none'; progressFill.style.width = '0%'; loadProjects(); loadStats(); }, 500);
+                } else {
+                    uploadError.textContent = 'Ошибка загрузки: ' + res.statusText;
+                    progressBar.style.display = 'none';
+                }
+            } catch (e) {
+                uploadError.textContent = 'Ошибка сети';
+                progressBar.style.display = 'none';
+            }
+        }
+        
+        // Toast notification
+        function showToast(msg, type) {
+            const toast = document.createElement('div');
+            toast.style.cssText = `position:fixed; bottom:20px; right:20px; background:rgba(0,0,0,0.95); color:white; padding:14px 24px; border-radius:8px; border:1px solid ${type==='success'?'#00c864':'#ff4444'}; z-index:100000; font-size:13px; box-shadow:0 4px 20px rgba(0,0,0,0.5);`;
+            toast.textContent = msg;
+            document.body.appendChild(toast);
+            setTimeout(() => toast.remove(), 4000);
+        }
+        
+        // Init
+        loadStats();
+        loadProjects();
+        setInterval(() => { loadProjects(); loadStats(); }, 5000); // Обновление каждые 5 сек
+    </script>
+</body>
+</html>

backend/static/review.html

@@ -0,0 +1,495 @@
+<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>Review Issues — Blueprint QC</title>
+    <style>
+        * { margin: 0; padding: 0; box-sizing: border-box; }
+        body {
+            font-family: 'Segoe UI', system-ui, sans-serif;
+            background: #0d0d1a;
+            color: #eee;
+            min-height: 100vh;
+        }
+        .container { max-width: 1400px; margin: 0 auto; padding: 20px; }
+        header {
+            display: flex;
+            justify-content: space-between;
+            align-items: center;
+            margin-bottom: 30px;
+            padding-bottom: 20px;
+            border-bottom: 2px solid #0f3460;
+        }
+        h1 { color: #e94560; }
+        .project-info { opacity: 0.7; font-size: 14px; }
+        
+        .filters {
+            display: flex;
+            gap: 15px;
+            margin-bottom: 25px;
+            flex-wrap: wrap;
+            align-items: center;
+        }
+        .filter-group {
+            display: flex;
+            gap: 8px;
+            align-items: center;
+        }
+        .filter-group label { font-size: 13px; opacity: 0.8; }
+        .filter-group select, .filter-group input {
+            background: #1a1a2e;
+            border: 1px solid #0f3460;
+            color: #eee;
+            padding: 6px 12px;
+            border-radius: 4px;
+            font-size: 13px;
+        }
+        .filter-btn {
+            background: #e94560;
+            border: none;
+            color: white;
+            padding: 6px 16px;
+            border-radius: 4px;
+            cursor: pointer;
+            font-size: 13px;
+        }
+        .filter-btn.secondary {
+            background: #1a1a2e;
+            border: 1px solid #0f3460;
+        }
+        
+        .progress-bar {
+            width: 100%;
+            height: 6px;
+            background: #1a1a2e;
+            border-radius: 3px;
+            margin-bottom: 25px;
+            overflow: hidden;
+        }
+        .progress-fill {
+            height: 100%;
+            background: linear-gradient(90deg, #e94560, #ffaa00);
+            transition: width 0.5s;
+        }
+        .progress-text {
+            font-size: 12px;
+            opacity: 0.6;
+            margin-bottom: 8px;
+        }
+        
+        .issues-grid {
+            display: grid;
+            grid-template-columns: repeat(auto-fill, minmax(350px, 1fr));
+            gap: 15px;
+        }
+        .issue-card {
+            background: #16213e;
+            border-left: 4px solid var(--card-color, #888);
+            border-radius: 8px;
+            padding: 15px;
+            transition: all 0.2s;
+            position: relative;
+        }
+        .issue-card:hover {
+            background: #1a2847;
+            transform: translateY(-2px);
+        }
+        .issue-card.reviewed-tp { border-left-color: #00c864; }
+        .issue-card.reviewed-fp { opacity: 0.6; }
+        .issue-card.reviewed-fp .issue-text { text-decoration: line-through; }
+        .issue-card.reviewed-ns { border-left-color: #888; }
+        
+        .issue-header {
+            display: flex;
+            justify-content: space-between;
+            align-items: center;
+            margin-bottom: 10px;
+        }
+        .issue-id {
+            font-size: 11px;
+            background: rgba(255,255,255,0.1);
+            padding: 2px 8px;
+            border-radius: 10px;
+        }
+        .issue-type {
+            font-size: 11px;
+            text-transform: uppercase;
+            opacity: 0.7;
+        }
+        .issue-severity {
+            font-size: 10px;
+            padding: 2px 8px;
+            border-radius: 10px;
+            font-weight: bold;
+        }
+        .severity-error { background: rgba(255,0,0,0.2); color: #ff4444; }
+        .severity-warning { background: rgba(255,170,0,0.2); color: #ffaa00; }
+        .severity-info { background: rgba(0,150,255,0.2); color: #44aaff; }
+        
+        .issue-text {
+            font-size: 13px;
+            line-height: 1.5;
+            margin-bottom: 10px;
+        }
+        .issue-meta {
+            font-size: 11px;
+            opacity: 0.6;
+            margin-bottom: 12px;
+        }
+        .issue-meta span { margin-right: 15px; }
+        
+        .feedback-bar {
+            display: flex;
+            gap: 8px;
+        }
+        .feedback-bar button {
+            flex: 1;
+            padding: 8px;
+            border: 1px solid #333;
+            background: #0d0d1a;
+            color: #ccc;
+            border-radius: 6px;
+            cursor: pointer;
+            font-size: 12px;
+            transition: all 0.2s;
+        }
+        .feedback-bar button:hover {
+            background: #1a1a2e;
+            color: #fff;
+        }
+        .feedback-bar button.tp:hover { background: rgba(0,200,100,0.2); border-color: #00c864; color: #00c864; }
+        .feedback-bar button.fp:hover { background: rgba(255,0,0,0.2); border-color: #ff4444; color: #ff4444; }
+        .feedback-bar button.ns:hover { background: rgba(150,150,150,0.2); border-color: #888; color: #888; }
+        .feedback-bar button:disabled { opacity: 0.5; cursor: not-allowed; }
+        
+        .reviewed-badge {
+            position: absolute;
+            top: 10px;
+            right: 10px;
+            font-size: 11px;
+            padding: 2px 8px;
+            border-radius: 10px;
+            font-weight: bold;
+        }
+        .badge-tp { background: rgba(0,200,100,0.2); color: #00c864; }
+        .badge-fp { background: rgba(255,0,0,0.2); color: #ff4444; }
+        .badge-ns { background: rgba(150,150,150,0.2); color: #888; }
+        
+        .empty-state {
+            text-align: center;
+            padding: 80px 20px;
+            opacity: 0.5;
+        }
+        .empty-state h2 { margin-bottom: 15px; }
+        
+        .viewer-link {
+            color: #44aaff;
+            text-decoration: none;
+            font-size: 12px;
+        }
+        .viewer-link:hover { text-decoration: underline; }
+        
+        #toast {
+            position: fixed;
+            bottom: 20px;
+            right: 20px;
+            background: rgba(0,0,0,0.95);
+            color: white;
+            padding: 14px 24px;
+            border-radius: 8px;
+            border: 1px solid #e94560;
+            z-index: 100000;
+            display: none;
+            font-size: 13px;
+            box-shadow: 0 4px 20px rgba(0,0,0,0.5);
+        }
+        
+        .stats-row {
+            display: flex;
+            gap: 20px;
+            margin-bottom: 20px;
+            flex-wrap: wrap;
+        }
+        .stat-pill {
+            background: #1a1a2e;
+            padding: 8px 16px;
+            border-radius: 20px;
+            font-size: 13px;
+        }
+        .stat-pill .num { font-weight: bold; color: #e94560; }
+    </style>
+</head>
+<body>
+    <div class="container">
+        <header>
+            <div>
+                <h1>Review Issues</h1>
+                <div class="project-info" id="projectInfo">Loading...</div>
+            </div>
+            <a href="/" style="color:#44aaff; text-decoration:none;">← Dashboard</a>
+        </header>
+        
+        <div class="stats-row" id="statsRow"></div>
+        
+        <div class="progress-text" id="progressText">0% reviewed</div>
+        <div class="progress-bar">
+            <div class="progress-fill" id="progressFill" style="width:0%"></div>
+        </div>
+        
+        <div class="filters">
+            <div class="filter-group">
+                <label>Project:</label>
+                <select id="projectSelect"><option value="">Select...</option></select>
+            </div>
+            <div class="filter-group">
+                <label>Severity:</label>
+                <select id="severityFilter">
+                    <option value="all">All</option>
+                    <option value="error">Error</option>
+                    <option value="warning">Warning</option>
+                    <option value="info">Info</option>
+                </select>
+            </div>
+            <div class="filter-group">
+                <label>Type:</label>
+                <select id="typeFilter"><option value="all">All</option></select>
+            </div>
+            <div class="filter-group">
+                <label>Status:</label>
+                <select id="statusFilter">
+                    <option value="all">All</option>
+                    <option value="unreviewed">Unreviewed</option>
+                    <option value="reviewed">Reviewed</option>
+                </select>
+            </div>
+            <button class="filter-btn" onclick="loadIssues()">Apply</button>
+            <button class="filter-btn secondary" onclick="exportTrainingData()">📥 Export Training Data</button>
+        </div>
+        
+        <div class="issues-grid" id="issuesGrid"></div>
+        
+        <div class="empty-state" id="emptyState" style="display:none;">
+            <h2>No issues found</h2>
+            <p>Try adjusting filters or upload a new PDF.</p>
+        </div>
+    </div>
+    
+    <div id="toast"></div>
+    
+    <script>
+        const API = '/api';
+        let currentProjectId = null;
+        let allIssues = [];
+        let allProjects = [];
+        
+        // Init
+        async function init() {
+            await loadProjects();
+            // Check URL param
+            const urlParams = new URLSearchParams(window.location.search);
+            const pid = urlParams.get('project');
+            if (pid) {
+                document.getElementById('projectSelect').value = pid;
+                await selectProject(pid);
+            }
+        }
+        
+        async function loadProjects() {
+            try {
+                const res = await fetch(`${API}/projects`);
+                allProjects = await res.json();
+                const select = document.getElementById('projectSelect');
+                select.innerHTML = '<option value="">Select project...</option>' + 
+                    allProjects.map(p => `<option value="${p.id}">${p.name || p.pdf_filename} (#${p.id})</option>`).join('');
+                select.addEventListener('change', (e) => selectProject(e.target.value));
+            } catch(e) { console.error(e); }
+        }
+        
+        async function selectProject(pid) {
+            if (!pid) return;
+            currentProjectId = pid;
+            const project = allProjects.find(p => p.id == pid);
+            if (project) {
+                document.getElementById('projectInfo').textContent = 
+                    `${project.name || project.pdf_filename} | Status: ${project.status}`;
+            }
+            await loadIssues();
+        }
+        
+        async function loadIssues() {
+            if (!currentProjectId) {
+                showToast('Select a project first', 'warning');
+                return;
+            }
+            
+            const severity = document.getElementById('severityFilter').value;
+            const typeFilter = document.getElementById('typeFilter').value;
+            const statusFilter = document.getElementById('statusFilter').value;
+            
+            try {
+                let url = `${API}/projects/${currentProjectId}/issues`;
+                const params = [];
+                if (severity !== 'all') params.push(`severity=${severity}`);
+                if (typeFilter !== 'all') params.push(`issue_type=${typeFilter}`);
+                if (statusFilter === 'reviewed') params.push(`has_feedback=true`);
+                if (statusFilter === 'unreviewed') params.push(`has_feedback=false`);
+                if (params.length) url += '?' + params.join('&');
+                
+                const res = await fetch(url);
+                allIssues = await res.json();
+                renderIssues();
+                updateStats();
+                updateTypeFilter();
+            } catch(e) { showToast('Error loading issues: ' + e.message, 'error'); }
+        }
+        
+        function renderIssues() {
+            const grid = document.getElementById('issuesGrid');
+            const empty = document.getElementById('emptyState');
+            
+            if (!allIssues.length) {
+                grid.innerHTML = '';
+                empty.style.display = 'block';
+                return;
+            }
+            empty.style.display = 'none';
+            
+            grid.innerHTML = allIssues.map(issue => {
+                const color = issue.severity === 'error' ? '#ff4444' : (issue.severity === 'warning' ? '#ffaa00' : '#44aaff');
+                const reviewedClass = issue.feedback ? 
+                    (issue.feedback.is_true_positive === true ? 'reviewed-tp' : 
+                     (issue.feedback.is_true_positive === false ? 'reviewed-fp' : 'reviewed-ns')) : '';
+                const badge = issue.feedback ? 
+                    (issue.feedback.is_true_positive === true ? '<span class="reviewed-badge badge-tp">✅ TP</span>' :
+                     (issue.feedback.is_true_positive === false ? '<span class="reviewed-badge badge-fp">❌ FP</span>' :
+                      '<span class="reviewed-badge badge-ns">🤷 NS</span>')) : '';
+                
+                return `
+                    <div class="issue-card ${reviewedClass}" style="--card-color:${color}" data-issue-id="${issue.id}">
+                        ${badge}
+                        <div class="issue-header">
+                            <span class="issue-id">#${issue.id}</span>
+                            <span class="issue-severity severity-${issue.severity}">${issue.severity}</span>
+                        </div>
+                        <div class="issue-type">${issue.issue_type}</div>
+                        <div class="issue-text">${issue.message}</div>
+                        <div class="issue-meta">
+                            <span>Page ${issue.page_number || '?'}</span>
+                            <span>Conf: ${issue.confidence ? issue.confidence.toFixed(2) : 'N/A'}</span>
+                            ${issue.dimension_text ? `<span>Dim: "${issue.dimension_text}"</span>` : ''}
+                            <a href="/viewer/${issue.project_id}/${issue.page_number || 1}" class="viewer-link" target="_blank">Open Viewer →</a>
+                        </div>
+                        <div class="feedback-bar">
+                            <button class="tp" onclick="submitFeedback(${issue.id}, true, this)" ${issue.feedback ? 'disabled' : ''}>✅ Real Issue</button>
+                            <button class="fp" onclick="submitFeedback(${issue.id}, false, this)" ${issue.feedback ? 'disabled' : ''}>❌ False Positive</button>
+                            <button class="ns" onclick="submitFeedback(${issue.id}, null, this)" ${issue.feedback ? 'disabled' : ''}>🤷 Not Sure</button>
+                        </div>
+                    </div>
+                `;
+            }).join('');
+        }
+        
+        async function submitFeedback(issueId, isTP, btn) {
+            const card = btn.closest('.issue-card');
+            const btns = card.querySelectorAll('.feedback-bar button');
+            btns.forEach(b => b.disabled = true);
+            btn.textContent = '...';
+            
+            try {
+                const res = await fetch(`${API}/feedback`, {
+                    method: 'POST',
+                    headers: {'Content-Type': 'application/json'},
+                    body: JSON.stringify({
+                        issue_id: issueId,
+                        is_true_positive: isTP,
+                        action_taken: isTP === true ? 'fixed' : (isTP === false ? 'ignored' : 'not_sure')
+                    })
+                });
+                
+                if (res.ok) {
+                    // Update the issue in memory
+                    const issue = allIssues.find(i => i.id === issueId);
+                    if (issue) {
+                        issue.feedback = { is_true_positive: isTP };
+                    }
+                    renderIssues();
+                    updateStats();
+                    showToast('Saved!', 'success');
+                } else {
+                    const err = await res.text();
+                    showToast('Error: ' + err, 'error');
+                    btns.forEach(b => b.disabled = false);
+                    btn.textContent = isTP === true ? '✅ Real Issue' : (isTP === false ? '❌ False Positive' : '🤷 Not Sure');
+                }
+            } catch(e) {
+                showToast('Network error: ' + e.message, 'error');
+                btns.forEach(b => b.disabled = false);
+                btn.textContent = isTP === true ? '✅ Real Issue' : (isTP === false ? '❌ False Positive' : '🤷 Not Sure');
+            }
+        }
+        
+        function updateStats() {
+            const total = allIssues.length;
+            const reviewed = allIssues.filter(i => i.feedback).length;
+            const tp = allIssues.filter(i => i.feedback && i.feedback.is_true_positive === true).length;
+            const fp = allIssues.filter(i => i.feedback && i.feedback.is_true_positive === false).length;
+            const pct = total ? Math.round((reviewed / total) * 100) : 0;
+            
+            document.getElementById('progressFill').style.width = pct + '%';
+            document.getElementById('progressText').textContent = 
+                `${pct}% reviewed (${reviewed}/${total}) | ✅ TP: ${tp} | ❌ FP: ${fp}`;
+            
+            document.getElementById('statsRow').innerHTML = `
+                <div class="stat-pill">Total: <span class="num">${total}</span></div>
+                <div class="stat-pill">Reviewed: <span class="num">${reviewed}</span></div>
+                <div class="stat-pill">True Positive: <span class="num" style="color:#00c864">${tp}</span></div>
+                <div class="stat-pill">False Positive: <span class="num" style="color:#ff4444">${fp}</span></div>
+            `;
+        }
+        
+        function updateTypeFilter() {
+            const types = [...new Set(allIssues.map(i => i.issue_type))];
+            const select = document.getElementById('typeFilter');
+            const current = select.value;
+            select.innerHTML = '<option value="all">All</option>' + 
+                types.map(t => `<option value="${t}">${t}</option>`).join('');
+            if (types.includes(current)) select.value = current;
+        }
+        
+        async function exportTrainingData() {
+            if (!currentProjectId) {
+                showToast('Select a project first', 'warning');
+                return;
+            }
+            try {
+                const res = await fetch(`${API}/training/export?project_id=${currentProjectId}&only_labeled=true`);
+                const data = await res.json();
+                
+                // Download as file
+                const blob = new Blob([JSON.stringify(data, null, 2)], {type: 'application/json'});
+                const url = URL.createObjectURL(blob);
+                const a = document.createElement('a');
+                a.href = url;
+                a.download = `training_data_project_${currentProjectId}_${new Date().toISOString().slice(0,10)}.json`;
+                a.click();
+                URL.revokeObjectURL(url);
+                
+                showToast(`Exported ${data.total_samples} samples!`, 'success');
+            } catch(e) {
+                showToast('Export error: ' + e.message, 'error');
+            }
+        }
+        
+        function showToast(msg, type) {
+            const toast = document.getElementById('toast');
+            toast.textContent = msg;
+            toast.style.borderColor = type === 'success' ? '#00c864' : (type === 'error' ? '#ff4444' : '#ffaa00');
+            toast.style.display = 'block';
+            setTimeout(() => toast.style.display = 'none', 3000);
+        }
+        
+        init();
+    </script>
+</body>
+</html>

compare_ocr.py

@@ -0,0 +1,160 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Визуализация сравнения: обычный OCR vs tiling OCR.
+Рисует bbox зелёным (только обычный), красным (только tiling), жёлтым (оба).
+"""
+
+import sys
+import json
+import re
+from pathlib import Path
+from PIL import Image, ImageDraw
+
+
+def load_ocr(path: Path):
+    """Загружает OCR lines из JSON."""
+    data = json.loads(path.read_text(encoding="utf-8"))
+    if "lines" in data:
+        return data["lines"]
+    if "pages" in data:
+        lines = []
+        for page in data["pages"]:
+            lines.extend(page.get("ocr_lines", []))
+        return lines
+    return []
+
+
+def bbox_center(box):
+    if isinstance(box[0], list):
+        xs = [p[0] for p in box]
+        ys = [p[1] for p in box]
+    else:
+        xs = [box[0], box[2]]
+        ys = [box[1], box[3]]
+    return sum(xs)/len(xs), sum(ys)/len(ys)
+
+
+def bbox_rect(box):
+    if isinstance(box[0], list):
+        xs = [p[0] for p in box]
+        ys = [p[1] for p in box]
+    else:
+        xs = [box[0], box[2]]
+        ys = [box[1], box[3]]
+    return min(xs), min(ys), max(xs), max(ys)
+
+
+def find_matches(text: str, list_b, iou_thresh=0.3):
+    """Находит ближайший совпадающий bbox в list_b по IoU и тексту."""
+    matches = []
+    for b in list_b:
+        if b["text"].strip() != text.strip():
+            continue
+        # IoU
+        ax1, ay1, ax2, ay2 = bbox_rect(a["bbox"] if 'a' in dir() else None)
+        # ... (упрощённо: сравниваем по центру)
+    return matches
+
+
+def visualize_comparison(png_path: Path, normal_ocr_path: Path, tiling_ocr_path: Path, out_path: Path):
+    """Рисует сравнение."""
+    img = Image.open(png_path)
+    draw = ImageDraw.Draw(img)
+    
+    normal = load_ocr(normal_ocr_path)
+    tiling = load_ocr(tiling_ocr_path)
+    
+    # Индексы для быстрого поиска
+    normal_by_text = {}
+    for n in normal:
+        txt = n["text"].strip()
+        if re.match(r'^\d+([,.]\d+)?$', txt):
+            normal_by_text.setdefault(txt, []).append(n)
+    
+    tiling_by_text = {}
+    for t in tiling:
+        txt = t["text"].strip()
+        if re.match(r'^\d+([,.]\d+)?$', txt):
+            tiling_by_text.setdefault(txt, []).append(t)
+    
+    # Классификация
+    only_normal = []      # зелёный
+    only_tiling = []      # красный
+    both = []             # жёлтый
+    
+    all_texts = set(normal_by_text.keys()) | set(tiling_by_text.keys())
+    
+    for txt in all_texts:
+        n_list = normal_by_text.get(txt, [])
+        t_list = tiling_by_text.get(txt, [])
+        
+        # Сопоставляем по минимальному расстоянию центров
+        used_t = set()
+        for n in n_list:
+            cx_n, cy_n = bbox_center(n["bbox"])
+            best = None
+            best_dist = float('inf')
+            for i, t in enumerate(t_list):
+                if i in used_t:
+                    continue
+                cx_t, cy_t = bbox_center(t["bbox"])
+                d = ((cx_n - cx_t)**2 + (cy_n - cy_t)**2)**0.5
+                if d < best_dist:
+                    best_dist = d
+                    best = i
+            
+            if best is not None and best_dist < 100:  # совпадение
+                both.append((n, t_list[best]))
+                used_t.add(best)
+            else:
+                only_normal.append(n)
+        
+        for i, t in enumerate(t_list):
+            if i not in used_t:
+                only_tiling.append(t)
+    
+    # Рисуем
+    for item in only_normal:
+        x1, y1, x2, y2 = bbox_rect(item["bbox"])
+        draw.rectangle([x1, y1, x2, y2], outline="green", width=3)
+    
+    for item in only_tiling:
+        x1, y1, x2, y2 = bbox_rect(item["bbox"])
+        draw.rectangle([x1, y1, x2, y2], outline="red", width=3)
+        cx, cy = bbox_center(item["bbox"])
+        draw.text((cx, cy-15), item["text"], fill="red")
+    
+    for n, t in both:
+        # Используем bbox из tiling (крупнее)
+        x1, y1, x2, y2 = bbox_rect(t["bbox"])
+        draw.rectangle([x1, y1, x2, y2], outline="yellow", width=2)
+    
+    img.save(out_path)
+    print(f"[OK] Сохранено: {out_path}")
+    print(f"  Только обычный (зелёный): {len(only_normal)}")
+    print(f"  Только tiling (красный): {len(only_tiling)}")
+    print(f"  Оба (жёлтый): {len(both)}")
+    
+    # Вывод новых чисел
+    print(f"\nНовые числа от tiling OCR:")
+    for item in sorted(only_tiling, key=lambda x: x["bbox"][0][1]):
+        cx, cy = bbox_center(item["bbox"])
+        print(f"  {item['text']:>10}  x={cx:>8.0f}  y={cy:>8.0f}")
+
+
+def main():
+    if len(sys.argv) < 4:
+        print("Usage: python compare_ocr.py <png> <normal_ocr.json> <tiling_ocr.json>")
+        sys.exit(1)
+    
+    png = Path(sys.argv[1])
+    normal = Path(sys.argv[2])
+    tiling = Path(sys.argv[3])
+    out = png.parent / f"{png.stem}_ocr_compare.png"
+    
+    visualize_comparison(png, normal, tiling, out)
+
+
+if __name__ == "__main__":
+    main()

dimension_extractor.py

@@ -0,0 +1,193 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Локальный детектор размеров на чертеже.
+
+Подход:
+1. Находим линии на PNG (Canny + HoughLinesP) — только горизонтальные/вертикальные
+2. Загружаем OCR результаты, фильтруем только числа (regex ^\d+([,.]\d+)?$)
+3. Для каждого числа проверяем: есть ли линия в радиусе 60px?
+4. Если да — считаем это размером
+5. Визуализируем результат
+
+Результат: dimensions.json + *_dims_detected.png
+"""
+
+import sys
+import json
+import math
+import re
+from pathlib import Path
+from typing import List, Dict, Tuple
+import cv2
+import numpy as np
+from PIL import Image, ImageDraw
+
+
+def find_numbers_with_context(ocr_path: Path, png_path: Path) -> Tuple[List[Dict], List[Dict]]:
+    """
+    Находит размеры через анализ контекста:
+    1. Берём все числа из OCR
+    2. Ищем "соседей" на той же горизонтали/вертикали (размерные цепочки)
+    3. Исключаем числа из таблиц (по bbox: справа на странице)
+    4. Проверяем пиксели между числами: есть ли линия?
+    """
+    print(f"[INFO] Обработка {png_path.name}...")
+    
+    ocr = json.loads(ocr_path.read_text(encoding="utf-8"))
+    img = cv2.imread(str(png_path), cv2.IMREAD_GRAYSCALE)
+    h, w = img.shape[:2]
+    
+    numbers = []
+    for page in ocr.get("pages", []):
+        for line_data in page.get("ocr_lines", []):
+            txt = line_data["text"].strip()
+            if not re.match(r'^\d+([,.]\d+)?$', txt):
+                continue
+            bbox = line_data.get("bbox")
+            if not bbox:
+                continue
+            if isinstance(bbox[0], list):
+                xs = [p[0] for p in bbox]
+                ys = [p[1] for p in bbox]
+            else:
+                xs = [bbox[0], bbox[2]]
+                ys = [bbox[1], bbox[3]]
+            cx = sum(xs) / len(xs)
+            cy = sum(ys) / len(ys)
+            # Определяем границы
+            x1, y1, x2, y2 = min(xs), min(ys), max(xs), max(ys)
+            numbers.append({
+                "text": txt,
+                "bbox": bbox,
+                "x1": x1, "y1": y1, "x2": x2, "y2": y2,
+                "cx": cx, "cy": cy,
+                "page": page["page_number"]
+            })
+    
+    print(f"[INFO] Всего чисел: {len(numbers)}")
+    
+    # Фильтр 1: исключаем числа из правой части таблиц (x > 0.5w и y > 0.1h)
+    # Это эвристика для данного чертежа
+    filtered = [n for n in numbers if not (n["x1"] > w * 0.55 and n["y1"] > h * 0.05)]
+    print(f"[INFO] После фильтра таблиц: {len(filtered)}")
+    
+    # Фильтр 2: ищем "пары" чисел на одной горизонтали (±15px по Y)
+    # Если между числами есть линия — это размерная цепочка
+    dimensions = []
+    used = set()
+    
+    for i, a in enumerate(filtered):
+        if i in used:
+            continue
+        # Ищем соседей на той же Y
+        neighbors = []
+        for j, b in enumerate(filtered):
+            if i == j or j in used:
+                continue
+            # Сравниваем Y (горизонтальная линия) или X (вертикальная)
+            dy = abs(a["cy"] - b["cy"])
+            dx = abs(a["cx"] - b["cx"])
+            if dy < 20 and dx > 30 and dx < 600:
+                # Проверяем, есть ли между ними тёмная линия
+                y_check = int((a["cy"] + b["cy"]) / 2)
+                x_start = min(int(a["cx"]), int(b["cx"]))
+                x_end = max(int(a["cx"]), int(b["cx"]))
+                line_px = img[y_check, x_start:x_end]
+                dark_ratio = np.sum(line_px < 200) / len(line_px) if len(line_px) > 0 else 0
+                if dark_ratio > 0.3:  # >30% тёмных пикселей
+                    neighbors.append((j, b, dx, "horizontal"))
+        
+        # Ищем вертикальных соседей
+        for j, b in enumerate(filtered):
+            if i == j or j in used:
+                continue
+            dx = abs(a["cx"] - b["cx"])
+            dy = abs(a["cy"] - b["cy"])
+            if dx < 20 and dy > 30 and dy < 600:
+                x_check = int((a["cx"] + b["cx"]) / 2)
+                y_start = min(int(a["cy"]), int(b["cy"]))
+                y_end = max(int(a["cy"]), int(b["cy"]))
+                line_px = img[y_start:y_end, x_check]
+                dark_ratio = np.sum(line_px < 200) / len(line_px) if len(line_px) > 0 else 0
+                if dark_ratio > 0.3:
+                    neighbors.append((j, b, dy, "vertical"))
+        
+        if neighbors:
+            # Берём ближайшего соседа
+            neighbors.sort(key=lambda x: x[2])
+            j, b, dist, orient = neighbors[0]
+            dimensions.append({
+                "text": a["text"],
+                "bbox": a["bbox"],
+                "neighbor_text": b["text"],
+                "distance": int(dist),
+                "orientation": orient,
+                "page": a["page"]
+            })
+            used.add(i)
+            used.add(j)
+    
+    # Одиночные числа — это скорее всего массы/количества из таблиц, игнорируем
+    
+    print(f"[INFO] Размеров найдено: {len(dimensions)}")
+    return numbers, dimensions
+
+
+def visualize(png_path: Path, all_numbers: List[Dict], dimensions: List[Dict], out_path: Path):
+    """Рисует визуализацию: размеры — красные, остальные числа — синие."""
+    img = Image.open(png_path)
+    draw = ImageDraw.Draw(img)
+    
+    # Все числа (синие)
+    dim_texts = {d["text"] for d in dimensions}
+    for num in all_numbers:
+        bbox = num["bbox"]
+        if isinstance(bbox[0], list):
+            pts = [(p[0], p[1]) for p in bbox]
+        else:
+            pts = [(bbox[0], bbox[1]), (bbox[2], bbox[1]), (bbox[2], bbox[3]), (bbox[0], bbox[3])]
+        color = "red" if num["text"] in dim_texts else "blue"
+        width = 3 if num["text"] in dim_texts else 1
+        draw.polygon(pts, outline=color, width=width)
+        if num["text"] in dim_texts:
+            x = min(p[0] for p in pts)
+            y = min(p[1] for p in pts)
+            draw.text((x, y-15), num["text"], fill="red")
+    
+    img.save(out_path)
+    print(f"[OK] Визуализация сохранена: {out_path}")
+
+
+def main():
+    if len(sys.argv) < 3:
+        print("Usage: python dimension_extractor.py <png> <ocr_json>")
+        sys.exit(1)
+    
+    png_path = Path(sys.argv[1])
+    ocr_path = Path(sys.argv[2])
+    out_json = png_path.parent / "dimensions.json"
+    out_png = png_path.parent / f"{png_path.stem}_dims_detected.png"
+    
+    all_numbers, dimensions = find_numbers_with_context(ocr_path, png_path)
+    
+    with open(out_json, "w", encoding="utf-8") as f:
+        json.dump({
+            "dimensions": dimensions,
+            "stats": {
+                "total_numbers": len(all_numbers),
+                "dimensions_found": len(dimensions)
+            }
+        }, f, ensure_ascii=False, indent=2)
+    print(f"[OK] Результаты сохранены: {out_json}")
+    
+    visualize(png_path, all_numbers, dimensions, out_png)
+    
+    print("\nНайденные размеры:")
+    for d in dimensions:
+        neighbor = f" → {d['neighbor_text']}" if d['neighbor_text'] else ""
+        print(f"  {d['text']}{neighbor}")
+
+
+if __name__ == "__main__":
+    main()

dimension_qc_checker.py

@@ -0,0 +1,349 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Quality Control проверка простановки размеров на чертежах.
+
+Ищет нарушения ЕСКД / ГОСТ 2.307 по координатам OCR + геометрии:
+  - Пересечение размерных линий
+  - Уплотнение (размеры слишком близко)
+  - Наложение размеров на текст/штриховку
+  - Низкая читаемость (low confidence)
+  - Размер внутри контура объекта
+  - Пропуски цепочек размеров
+
+Использование:
+    python dimension_qc_checker.py <output_folder>
+
+Результат: <folder>/dimension_qc_report.json + текстовый отчёт
+"""
+
+import sys
+import json
+import re
+from pathlib import Path
+from typing import List, Dict, Tuple
+from dataclasses import dataclass
+
+# ------------------------------------------------------------------
+# QC правила
+# ------------------------------------------------------------------
+MIN_CONFIDENCE = 0.65        # ниже — подозрение на нечитаемость
+MIN_BBOX_OVERLAP = 0.15      # минимальное пересечение bbox для флага
+MIN_DIMENSION_SPACING = 8   # мин. пикселей между размерами (px @ 300 DPI ~ 0.7 мм)
+MAX_DIMENSION_CHAIN_GAP = 50  # макс. зазор между размерами в одной цепочке
+
+@dataclass
+class TextItem:
+    text: str
+    confidence: float
+    bbox: list  # [x1, y1, x2, y2]
+    page: int
+    is_dimension: bool = False
+    
+    @property
+    def x1(self) -> float:
+        return min(p[0] for p in self.bbox)
+    
+    @property
+    def y1(self) -> float:
+        return min(p[1] for p in self.bbox)
+    
+    @property
+    def x2(self) -> float:
+        return max(p[0] for p in self.bbox)
+    
+    @property
+    def y2(self) -> float:
+        return max(p[1] for p in self.bbox)
+    
+    @property
+    def width(self) -> float:
+        return self.x2 - self.x1
+    
+    @property
+    def height(self) -> float:
+        return self.y2 - self.y1
+    
+    @property
+    def center_x(self) -> float:
+        return (self.x1 + self.x2) / 2
+    
+    @property
+    def center_y(self) -> float:
+        return (self.y1 + self.y2) / 2
+    
+    @property
+    def area(self) -> float:
+        return self.width * self.height
+
+
+def is_dimension_text(text: str) -> bool:
+    """Эвристика: похоже ли на размер."""
+    text = text.strip().replace(' ', '').replace(',', '.')
+    # Чистые числа 50-50000
+    if re.match(r'^\d{2,5}(\.\d{1,2})?$', text):
+        num = float(text)
+        return 50 <= num <= 50000
+    # Числа с единицами
+    if re.match(r'^\d{2,5}(\.\d{1,2})?[мmмм]?[мm]?$', text, re.I):
+        return True
+    return False
+
+
+def bbox_overlap(a: TextItem, b: TextItem) -> float:
+    """IOU (Intersection over Union) двух bbox."""
+    x1 = max(a.x1, b.x1)
+    y1 = max(a.y1, b.y1)
+    x2 = min(a.x2, b.x2)
+    y2 = min(a.y2, b.y2)
+    if x2 <= x1 or y2 <= y1:
+        return 0.0
+    inter = (x2 - x1) * (y2 - y1)
+    union = a.area + b.area - inter
+    return inter / union if union > 0 else 0.0
+
+
+def bbox_distance(a: TextItem, b: TextItem) -> float:
+    """Расстояние между центрами bbox."""
+    import math
+    return math.sqrt((a.center_x - b.center_x)**2 + (a.center_y - b.center_y)**2)
+
+
+def analyze_dimensions(items: List[TextItem]) -> List[Dict]:
+    """Находит проблемы с размерами."""
+    issues = []
+    dims = [it for it in items if it.is_dimension]
+    non_dims = [it for it in items if not it.is_dimension]
+    
+    # --- 1. Низкий confidence (подозрение на нечитаемость) ---
+    for d in dims:
+        if d.confidence < MIN_CONFIDENCE:
+            issues.append({
+                "type": "LOW_CONFIDENCE_DIMENSION",
+                "severity": "warning",
+                "page": d.page,
+                "text": d.text,
+                "confidence": d.confidence,
+                "bbox": d.bbox,
+                "message": f"Размер '{d.text}' имеет низкую уверенность OCR ({d.confidence:.2f}). Возможно, плохо читается на чертеже. Рекомендуется увеличить или перепроставить.",
+            })
+    
+    # --- 2. Пересечение размеров (наложение) ---
+    for i, d1 in enumerate(dims):
+        for d2 in dims[i+1:]:
+            if d1.page != d2.page:
+                continue
+            overlap = bbox_overlap(d1, d2)
+            if overlap > MIN_BBOX_OVERLAP:
+                issues.append({
+                    "type": "DIMENSION_OVERLAP",
+                    "severity": "error",
+                    "page": d1.page,
+                    "text1": d1.text,
+                    "text2": d2.text,
+                    "overlap": overlap,
+                    "bbox1": d1.bbox,
+                    "bbox2": d2.bbox,
+                    "message": f"Размеры '{d1.text}' и '{d2.text}' пересекаются ({overlap:.0%} наложения). Нарушение ЕСКД: размерные числа не должны пересекать друг друга.",
+                })
+    
+    # --- 3. Наложение размеров на другой текст/штриховку ---
+    for d in dims:
+        for nd in non_dims:
+            if d.page != nd.page:
+                continue
+            # Проверяем, находится ли центр размера внутри bbox текста
+            if (nd.x1 < d.center_x < nd.x2 and 
+                nd.y1 < d.center_y < nd.y2):
+                issues.append({
+                    "type": "DIMENSION_ON_TEXT",
+                    "severity": "error",
+                    "page": d.page,
+                    "dimension": d.text,
+                    "overlapped_text": nd.text,
+                    "bbox_dim": d.bbox,
+                    "bbox_text": nd.bbox,
+                    "message": f"Размер '{d.text}' наложен на текст '{nd.text}'. Нарушение ЕСКД: размерные числа не должны пересекать линии контура или другие надписи.",
+                })
+    
+    # --- 4. "Уплотнение" размеров (цепочка без отступов) ---
+    # Группируем по страницам и по горизонтальным линиям (похожие Y)
+    from collections import defaultdict
+    page_dims = defaultdict(list)
+    for d in dims:
+        page_dims[d.page].append(d)
+    
+    for page, page_items in page_dims.items():
+        # Сортируем по Y
+        page_items.sort(key=lambda x: x.center_y)
+        chains = []
+        current_chain = [page_items[0]] if page_items else []
+        
+        for d in page_items[1:]:
+            prev = current_chain[-1]
+            # Если Y близко — считаем одной цепочкой
+            if abs(d.center_y - prev.center_y) < 25:  # допуск по вертикали
+                current_chain.append(d)
+            else:
+                if len(current_chain) >= 3:
+                    chains.append(current_chain)
+                current_chain = [d]
+        if len(current_chain) >= 3:
+            chains.append(current_chain)
+        
+        # Проверяем цепочки на уплотнение
+        for chain in chains:
+            chain.sort(key=lambda x: x.center_x)
+            for i in range(1, len(chain)):
+                gap = chain[i].center_x - chain[i-1].center_x
+                # Если размеры ближе чем ~2× их высота — это уплотнение
+                avg_height = (chain[i].height + chain[i-1].height) / 2
+                if gap < avg_height * 1.5:
+                    issues.append({
+                        "type": "DIMENSION_CROWDING",
+                        "severity": "warning",
+                        "page": page,
+                        "text1": chain[i-1].text,
+                        "text2": chain[i].text,
+                        "gap_px": gap,
+                        "bbox1": chain[i-1].bbox,
+                        "bbox2": chain[i].bbox,
+                        "message": f"Размеры '{chain[i-1].text}' и '{chain[i].text}' расположены слишком близко (зазор {gap:.0f} px). Возможно 'уплотнение' — размерные линии не отступают друг от друга. Рекомендуется разнести.",
+                    })
+    
+    # --- 5. Пропуски в цепочке размеров (gaps) ---
+    # Если в цепочке размеров есть большой зазор без размера — возможно пропущен
+    for page, page_items in page_dims.items():
+        page_items.sort(key=lambda x: x.center_x)
+        for i in range(1, len(page_items)):
+            gap = page_items[i].center_x - page_items[i-1].center_x
+            avg_width = (page_items[i].width + page_items[i-1].width) / 2
+            # Если зазор в 4+ раза больше среднего размера — подозрительно
+            if gap > avg_width * 4:
+                issues.append({
+                    "type": "DIMENSION_CHAIN_GAP",
+                    "severity": "info",
+                    "page": page,
+                    "left": page_items[i-1].text,
+                    "right": page_items[i].text,
+                    "gap_px": gap,
+                    "message": f"Между размерами '{page_items[i-1].text}' и '{page_items[i].text}' большой зазор ({gap:.0f} px). Возможно, пропущен промежуточный размер в цепочке.",
+                })
+    
+    return issues
+
+
+def validate_folder(folder: Path):
+    ocr_path = folder / "full_ocr_results.json"
+    if not ocr_path.exists():
+        print(f"[ERR] Не найден {ocr_path}")
+        sys.exit(1)
+    
+    data = json.loads(ocr_path.read_text(encoding="utf-8"))
+    pages = data["pages"]
+    
+    all_items = []
+    for page in pages:
+        page_num = page["page_number"]
+        for entry in page.get("ocr_lines", []):
+            item = TextItem(
+                text=entry["text"],
+                confidence=entry.get("confidence", 0),
+                bbox=entry.get("bbox", [0,0,0,0]),
+                page=page_num,
+                is_dimension=is_dimension_text(entry["text"]),
+            )
+            all_items.append(item)
+    
+    dimensions = [it for it in all_items if it.is_dimension]
+    
+    print(f"[INFO] Всего элементов: {len(all_items)}")
+    print(f"[INFO] Размеров найдено: {len(dimensions)}")
+    print(f"[INFO] Проверка...\n")
+    
+    issues = analyze_dimensions(all_items)
+    
+    # Группировка по серьёзности
+    errors = [i for i in issues if i["severity"] == "error"]
+    warnings = [i for i in issues if i["severity"] == "warning"]
+    infos = [i for i in issues if i["severity"] == "info"]
+    
+    print("=" * 70)
+    print("ОШИБКИ (требуют переделки)")
+    print("=" * 70)
+    if errors:
+        for i, iss in enumerate(errors, 1):
+            print(f"\n[{i}] Стр.{iss['page']}: {iss['type']}")
+            print(f"    {iss['message']}")
+    else:
+        print("\n✅ Критических ошибок не найдено")
+    
+    print("\n" + "=" * 70)
+    print("ПРЕДУПРЕЖДЕНИЯ (рекомендуется исправить)")
+    print("=" * 70)
+    if warnings:
+        for i, iss in enumerate(warnings[:20], 1):
+            print(f"\n[{i}] Стр.{iss['page']}: {iss['type']}")
+            print(f"    {iss['message']}")
+        if len(warnings) > 20:
+            print(f"\n... и ещё {len(warnings) - 20} предупреждений")
+    else:
+        print("\n✅ Предупреждений не найдено")
+    
+    if infos:
+        print(f"\nℹ️ Информационных замечаний: {len(infos)}")
+    
+    # Сохранение JSON
+    report = {
+        "summary": {
+            "total_items": len(all_items),
+            "dimensions_found": len(dimensions),
+            "errors": len(errors),
+            "warnings": len(warnings),
+            "infos": len(infos),
+        },
+        "errors": errors,
+        "warnings": warnings,
+        "infos": infos,
+    }
+    out_path = folder / "dimension_qc_report.json"
+    with open(out_path, "w", encoding="utf-8") as f:
+        json.dump(report, f, ensure_ascii=False, indent=2)
+    print(f"\n[INFO] Отчёт сохранён: {out_path}")
+    
+    # Сгенерировать markdown-замечания для проектировщика
+    md_lines = ["# Замечания по простановке размеров\n"]
+    md_lines.append(f"**Документ:** {folder.name}\n")
+    md_lines.append(f"**Всего размеров:** {len(dimensions)}\n")
+    md_lines.append(f"**Ошибок:** {len(errors)} | **Предупреждений:** {len(warnings)}\n\n")
+    
+    if errors:
+        md_lines.append("## Ошибки (обязательно к исправлению)\n\n")
+        for i, iss in enumerate(errors, 1):
+            md_lines.append(f"### {i}. {iss['type']} (стр. {iss['page']})\n\n")
+            md_lines.append(f"{iss['message']}\n\n")
+            if 'bbox1' in iss:
+                md_lines.append(f"- Координаты 1: `{iss['bbox1']}`\n")
+            if 'bbox2' in iss:
+                md_lines.append(f"- Координаты 2: `{iss['bbox2']}`\n")
+            md_lines.append("\n")
+    
+    if warnings:
+        md_lines.append("## Предупреждения (рекомендуется исправить)\n\n")
+        for i, iss in enumerate(warnings[:10], 1):
+            md_lines.append(f"### {i}. {iss['type']} (стр. {iss['page']})\n\n")
+            md_lines.append(f"{iss['message']}\n\n")
+    
+    md_path = folder / "dimension_qc_remarks.md"
+    with open(md_path, "w", encoding="utf-8") as f:
+        f.writelines(md_lines)
+    print(f"[INFO] Замечания для проектировщика: {md_path}")
+
+
+def main():
+    folder = Path(sys.argv[1]) if len(sys.argv) > 1 else Path("output_123")
+    validate_folder(folder)
+
+
+if __name__ == "__main__":
+    main()

generate_dzi.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Генератор Deep Zoom Image (DZI) тайлов из PNG для быстрого просмотра больших чертежей.
+
+Использует PIL — не требует внешних зависимостей.
+
+Использование:
+    python generate_dzi.py <png_file> [--tile-size 256] [--format png]
+    
+Результат:
+    <png_stem>.dzi          — XML-дескриптор
+    <png_stem>_files/       — папка с тайлами level/col_row.png
+"""
+
+import sys
+import os
+import math
+from pathlib import Path
+from PIL import Image
+
+
+def generate_dzi(png_path: Path, tile_size: int = 256, fmt: str = "png"):
+    """Генерирует DZI тайлы из PNG."""
+    png_path = Path(png_path)
+    if not png_path.exists():
+        print(f"[ERR] Файл не найден: {png_path}")
+        sys.exit(1)
+    
+    base = png_path.stem
+    files_dir = png_path.parent / f"{base}_files"
+    files_dir.mkdir(exist_ok=True)
+    
+    print(f"[INFO] Открываем {png_path}...")
+    img = Image.open(png_path)
+    orig_w, orig_h = img.size
+    print(f"[INFO] Размер: {orig_w}x{orig_h}")
+    
+    # DZI uses power-of-2 levels, starting from 1x1 at level 0
+    max_dim = max(orig_w, orig_h)
+    max_level = math.ceil(math.log2(max_dim))
+    
+    print(f"[INFO] Уровней: {max_level + 1} (0..{max_level})")
+    
+    # Process from largest to smallest (build pyramid top-down)
+    current = img.convert("RGBA" if fmt == "png" else "RGB")
+    
+    for level in range(max_level, -1, -1):
+        # Calculate size at this level
+        scale = 2 ** (max_level - level)
+        level_w = math.ceil(orig_w / scale)
+        level_h = math.ceil(orig_h / scale)
+        
+        # Resize current image to level size
+        if current.size != (level_w, level_h):
+            current = current.resize((level_w, level_h), Image.LANCZOS)
+        
+        # Save tiles
+        cols = math.ceil(level_w / tile_size)
+        rows = math.ceil(level_h / tile_size)
+        
+        level_dir = files_dir / str(level)
+        level_dir.mkdir(exist_ok=True)
+        
+        for row in range(rows):
+            for col in range(cols):
+                x = col * tile_size
+                y = row * tile_size
+                w = min(tile_size, level_w - x)
+                h = min(tile_size, level_h - y)
+                
+                tile = current.crop((x, y, x + w, y + h))
+                tile_path = level_dir / f"{col}_{row}.{fmt}"
+                
+                if fmt == "png":
+                    tile.save(tile_path, "PNG", compress_level=3)
+                else:
+                    tile.save(tile_path, "JPEG", quality=85)
+        
+        print(f"  Level {level}: {level_w}x{level_h}, {cols}x{rows} tiles")
+    
+    # Write DZI descriptor
+    dzi_path = png_path.parent / f"{base}.dzi"
+    dzi_xml = f'''<?xml version="1.0" encoding="UTF-8"?>
+<Image xmlns="http://schemas.microsoft.com/deepzoom/2008"
+       Format="{fmt}"
+       Overlap="0"
+       TileSize="{tile_size}">
+    <Size Width="{orig_w}"
+          Height="{orig_h}"/>
+</Image>'''
+    dzi_path.write_text(dzi_xml, encoding="utf-8")
+    
+    print(f"[OK] DZI создан: {dzi_path}")
+    print(f"     Тайлы: {files_dir}")
+    print(f"     Уровней: {max_level + 1}, TileSize: {tile_size}")
+    
+    return dzi_path, files_dir
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: python generate_dzi.py <png_file> [tile_size] [format]")
+        sys.exit(1)
+    
+    png = Path(sys.argv[1])
+    tile_size = int(sys.argv[2]) if len(sys.argv) > 2 else 256
+    fmt = sys.argv[3] if len(sys.argv) > 3 else "png"
+    
+    generate_dzi(png, tile_size, fmt)
+
+
+if __name__ == "__main__":
+    main()

generate_web_viewer.py

@@ -0,0 +1,857 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Генерация HTML-viewer с OpenSeadragon + DZI для просмотра замечаний QC прямо на чертеже.
+
+Использование:
+    python generate_web_viewer.py <output_folder> [--page N] [--dzi]
+    
+Требует предварительного запуска generate_dzi.py для создания тайлов.
+"""
+
+import sys
+import json
+import base64
+import shutil
+from pathlib import Path
+
+
+def generate_html(folder: Path, target_page: int = None, use_dzi: bool = True, 
+                   issue_db_ids: list = None, total_pages: int = None, 
+                   project_id: int = None, api_base: str = None):
+    """Генерирует HTML-viewer с OpenSeadragon + overlay."""
+    
+    qc_path = folder / "dimension_qc_report.json"
+    if not qc_path.exists():
+        print(f"[ERR] Сначала запустите dimension_qc_checker.py")
+        sys.exit(1)
+    
+    qc = json.loads(qc_path.read_text(encoding="utf-8"))
+    
+    # Добавить VLM QC issues если есть
+    vlm_path = folder / "vlm_qc_report.json"
+    if vlm_path.exists():
+        vlm_qc = json.loads(vlm_path.read_text(encoding="utf-8"))
+        for severity in ["errors", "warnings", "infos"]:
+            qc.setdefault(severity, [])
+            qc[severity].extend(vlm_qc.get(severity, []))
+    
+    ocr = json.loads((folder / "full_ocr_results.json").read_text(encoding="utf-8"))
+    
+    page_counts = {}
+    for severity in ["errors", "warnings", "infos"]:
+        for item in qc.get(severity, []):
+            p = item.get("page", 1)
+            page_counts[p] = page_counts.get(p, 0) + 1
+    
+    if target_page is None:
+        target_page = max(page_counts, key=page_counts.get) if page_counts else 2
+    
+    print(f"[INFO] Генерация viewer для страницы {target_page}")
+    
+    # Определить пути к изображению
+    png_path = folder / f"page_{target_page:03d}.png"
+    dzi_path = folder / f"page_{target_page:03d}.dzi"
+    
+    if use_dzi and not dzi_path.exists():
+        print(f"[WARN] DZI не найден: {dzi_path}")
+        print(f"       Запустите: python generate_dzi.py {png_path}")
+        use_dzi = False
+    
+    # Получить размеры PNG
+    from PIL import Image
+    with Image.open(png_path) as img:
+        img_width, img_height = img.size
+    
+    # Собрать элементы страницы
+    page_items = []
+    for page in ocr["pages"]:
+        if page["page_number"] == target_page:
+            for line in page.get("ocr_lines", []):
+                page_items.append(line)
+            break
+    
+    # Собрать проблемы
+    issues = []
+    colors = {"error": "#ff0000", "warning": "#ffaa00", "info": "#0099ff"}
+    for severity in ["errors", "warnings", "infos"]:
+        for item in qc.get(severity, []):
+            if item["page"] == target_page:
+                bboxes = []
+                for key in ["bbox1", "bbox2", "bbox", "bbox_dim"]:
+                    if key in item:
+                        bboxes.append(item[key])
+                
+                if not bboxes and "text" in item:
+                    for line in page_items:
+                        if line["text"] == item["text"]:
+                            bboxes.append(line["bbox"])
+                            break
+                
+                issues.append({
+                    "type": item["type"],
+                    "message": item["message"],
+                    "severity": item["severity"],
+                    "color": colors.get(item["severity"], "#999"),
+                    "bboxes": bboxes,
+                    "source": item.get("source", "rules"),
+                })
+    
+    # Подготовить overlay-данные для JS
+    overlay_data = []
+    issue_counter = 0
+    for issue in issues:
+        issue_counter += 1
+        for bbox in issue["bboxes"]:
+            if isinstance(bbox[0], list):
+                xs = [p[0] for p in bbox]
+                ys = [p[1] for p in bbox]
+                x1, y1, x2, y2 = min(xs), min(ys), max(xs), max(ys)
+            else:
+                x1, y1, x2, y2 = bbox[0], bbox[1], bbox[2], bbox[3]
+            
+            # Convert pixel to viewport (0-1)
+            vx = x1 / img_width
+            vy = y1 / img_height
+            vw = (x2 - x1) / img_width
+            vh = (y2 - y1) / img_height
+            
+            overlay_data.append({
+                "id": f"issue-{issue_counter}",
+                "x": round(vx, 6),
+                "y": round(vy, 6),
+                "w": round(vw, 6),
+                "h": round(vh, 6),
+                "color": issue["color"],
+                "message": issue["message"],
+                "type": issue["type"],
+                "severity": issue["severity"],
+                "num": issue_counter,
+            })
+    
+    # JSON для вставки в JS
+    overlays_json = json.dumps(overlay_data, ensure_ascii=False)
+    
+    # DZI или прямой PNG
+    if use_dzi:
+        # Inline DZI to avoid CORS issues with file:// protocol
+        dzi_xml = (folder / f"page_{target_page:03d}.dzi").read_text(encoding="utf-8")
+        # Parse key values
+        import re
+        w = re.search(r'Width="(\d+)"', dzi_xml).group(1)
+        h = re.search(r'Height="(\d+)"', dzi_xml).group(1)
+        ts = re.search(r'TileSize="(\d+)"', dzi_xml).group(1)
+        fmt = re.search(r'Format="(\w+)"', dzi_xml).group(1)
+        tile_source = f"""{{
+            Image: {{
+                xmlns: "http://schemas.microsoft.com/deepzoom/2008",
+                Url: "./page_{target_page:03d}_files/",
+                Format: "{fmt}",
+                Overlap: "0",
+                TileSize: "{ts}",
+                Size: {{ Width: "{w}", Height: "{h}" }}
+            }}
+        }}"""
+        dzi_note = ""
+    else:
+        png_data = base64.b64encode(png_path.read_bytes()).decode()
+        tile_source = f"{{ type: 'image', url: 'data:image/png;base64,{png_data}' }}"
+        dzi_note = "<div style='color:#ffaa00; padding:10px;'>⚠ DZI не найден — используется прямое PNG (медленно для больших чертежей)</div>"
+    
+    # Генерация HTML
+    html = f'''<!DOCTYPE html>
+<html lang="ru">
+<head>
+    <meta charset="UTF-8">
+    <title>QC Viewer — Страница {target_page}</title>
+    <script src="https://cdn.jsdelivr.net/npm/openseadragon@4.1/build/openseadragon/openseadragon.min.js"></script>
+    <style>
+        * {{ box-sizing: border-box; margin: 0; padding: 0; }}
+        body {{
+            font-family: 'Segoe UI', system-ui, sans-serif;
+            background: #0d0d1a;
+            color: #eee;
+            overflow: hidden;
+            height: 100vh;
+        }}
+        .layout {{
+            display: flex;
+            height: 100vh;
+        }}
+        .viewer-area {{
+            flex: 1;
+            position: relative;
+            background: #080810;
+        }}
+        #openseadragon {{
+            width: 100%;
+            height: 100%;
+        }}
+        .overlay-rect {{
+            border: 2px solid var(--oc-color);
+            background: var(--oc-color);
+            opacity: 0.15;
+            cursor: pointer;
+            transition: opacity 0.2s, border-width 0.2s;
+            position: relative;
+        }}
+        .overlay-rect:hover {{
+            opacity: 0.5;
+            border-width: 3px;
+            z-index: 100;
+        }}
+        .overlay-number {{
+            position: absolute;
+            top: -18px;
+            left: 0;
+            color: var(--oc-color);
+            font-size: 13px;
+            font-weight: bold;
+            text-shadow: 1px 1px 2px rgba(0,0,0,0.8);
+            pointer-events: none;
+        }}
+        .sidebar {{
+            width: 400px;
+            background: #16213e;
+            border-left: 2px solid #0f3460;
+            display: flex;
+            flex-direction: column;
+            overflow: hidden;
+        }}
+        .sidebar-header {{
+            padding: 20px;
+            border-bottom: 2px solid #0f3460;
+        }}
+        .sidebar-header h2 {{
+            margin: 0 0 15px 0;
+            color: #e94560;
+            font-size: 18px;
+        }}
+        .stats {{
+            display: flex;
+            gap: 10px;
+        }}
+        .stat {{
+            flex: 1;
+            background: #1a1a2e;
+            padding: 10px;
+            border-radius: 6px;
+            text-align: center;
+        }}
+        .stat-value {{
+            font-size: 22px;
+            font-weight: bold;
+        }}
+        .stat.error .stat-value {{ color: #ff4444; }}
+        .stat.warning .stat-value {{ color: #ffaa00; }}
+        .stat.info .stat-value {{ color: #44aaff; }}
+        .stat-label {{
+            font-size: 11px;
+            opacity: 0.7;
+            margin-top: 4px;
+        }}
+        .legend {{
+            display: flex;
+            gap: 15px;
+            margin-top: 12px;
+            font-size: 12px;
+        }}
+        .legend-item {{
+            display: flex;
+            align-items: center;
+            gap: 6px;
+        }}
+        .legend-dot {{
+            width: 12px;
+            height: 12px;
+            border-radius: 3px;
+        }}
+        .issue-list {{
+            flex: 1;
+            overflow-y: auto;
+            padding: 15px;
+        }}
+        .issue-card {{
+            background: #1a1a2e;
+            border-left: 4px solid var(--card-color);
+            padding: 12px;
+            margin-bottom: 10px;
+            border-radius: 4px;
+            cursor: pointer;
+            transition: all 0.2s;
+        }}
+        .issue-card:hover {{
+            background: #252540;
+            transform: translateX(4px);
+        }}
+        .issue-card.active {{
+            background: #2a2a50;
+            box-shadow: 0 0 0 2px var(--card-color);
+        }}
+        .issue-num {{
+            display: inline-block;
+            background: var(--card-color);
+            color: #000;
+            font-size: 11px;
+            font-weight: bold;
+            padding: 2px 8px;
+            border-radius: 10px;
+            margin-right: 8px;
+        }}
+        .issue-type {{
+            font-size: 11px;
+            text-transform: uppercase;
+            opacity: 0.6;
+            margin-bottom: 6px;
+        }}
+        .issue-message {{
+            font-size: 13px;
+            line-height: 1.4;
+        }}
+        .issue-severity {{
+            font-size: 10px;
+            opacity: 0.5;
+            margin-top: 6px;
+            text-transform: uppercase;
+        }}
+        .feedback-bar {{
+            display: flex;
+            gap: 6px;
+            margin-top: 10px;
+            padding-top: 10px;
+            border-top: 1px solid rgba(255,255,255,0.1);
+        }}
+        .feedback-bar button {{
+            flex: 1;
+            padding: 4px 8px;
+            border: none;
+            border-radius: 4px;
+            font-size: 11px;
+            cursor: pointer;
+            transition: all 0.2s;
+            background: #1a1a2e;
+            color: #ccc;
+            border: 1px solid #333;
+        }}
+        .feedback-bar button:hover {{
+            background: #252540;
+            color: #fff;
+        }}
+        .feedback-bar button.tp:hover {{ background: rgba(0,200,100,0.2); border-color: #00c864; color: #00c864; }}
+        .feedback-bar button.fp:hover {{ background: rgba(255,0,0,0.2); border-color: #ff4444; color: #ff4444; }}
+        .feedback-bar button.ns:hover {{ background: rgba(150,150,150,0.2); border-color: #888; color: #888; }}
+        .issue-card.feedback-tp {{ border-left-color: #00c864 !important; box-shadow: 0 0 0 2px rgba(0,200,100,0.3); }}
+        .issue-card.feedback-fp {{ opacity: 0.5; }}
+        .issue-card.feedback-fp .issue-message {{ text-decoration: line-through; }}
+        .feedback-status {{
+            font-size: 11px;
+            margin-top: 6px;
+            font-weight: bold;
+        }}
+        .feedback-status.tp {{ color: #00c864; }}
+        .feedback-status.fp {{ color: #ff4444; }}
+        .feedback-status.ns {{ color: #888; }}
+        #toast {{
+            position: fixed;
+            bottom: 20px;
+            right: 20px;
+            background: rgba(0,0,0,0.9);
+            color: white;
+            padding: 12px 20px;
+            border-radius: 8px;
+            border: 1px solid #e94560;
+            z-index: 100000;
+            display: none;
+            font-size: 13px;
+            box-shadow: 0 4px 20px rgba(0,0,0,0.5);
+        }}
+        #tooltip {{
+            position: fixed;
+            background: rgba(0,0,0,0.92);
+            color: white;
+            padding: 12px 16px;
+            border-radius: 8px;
+            font-size: 13px;
+            max-width: 350px;
+            pointer-events: none;
+            z-index: 10000;
+            display: none;
+            border: 1px solid #e94560;
+            box-shadow: 0 4px 20px rgba(0,0,0,0.5);
+        }}
+        .controls {{
+            position: absolute;
+            top: 15px;
+            left: 15px;
+            z-index: 1000;
+            background: rgba(22,33,62,0.95);
+            padding: 10px;
+            border-radius: 8px;
+            border: 1px solid #0f3460;
+        }}
+        .controls button {{
+            background: #1a1a2e;
+            border: 1px solid #0f3460;
+            color: #eee;
+            padding: 6px 12px;
+            border-radius: 4px;
+            cursor: pointer;
+            font-size: 12px;
+            margin-right: 5px;
+        }}
+        .controls button:hover {{
+            background: #252540;
+        }}
+        .controls button.active {{
+            background: #e94560;
+            border-color: #e94560;
+        }}
+        .page-nav {{
+            position: absolute;
+            bottom: 20px;
+            left: 50%;
+            transform: translateX(-50%);
+            z-index: 1000;
+            background: rgba(22,33,62,0.95);
+            padding: 10px 20px;
+            border-radius: 30px;
+            border: 1px solid #0f3460;
+            display: flex;
+            align-items: center;
+            gap: 15px;
+            box-shadow: 0 4px 20px rgba(0,0,0,0.5);
+        }}
+        .page-nav button {{
+            background: #1a1a2e;
+            border: 1px solid #0f3460;
+            color: #eee;
+            padding: 8px 16px;
+            border-radius: 6px;
+            cursor: pointer;
+            font-size: 13px;
+        }}
+        .page-nav button:hover:not(:disabled) {{
+            background: #252540;
+        }}
+        .page-nav button:disabled {{
+            opacity: 0.3;
+            cursor: not-allowed;
+        }}
+        .page-nav select {{
+            background: #1a1a2e;
+            border: 1px solid #0f3460;
+            color: #eee;
+            padding: 8px 12px;
+            border-radius: 6px;
+            font-size: 13px;
+            cursor: pointer;
+        }}
+        .page-nav .page-info {{
+            font-size: 13px;
+            opacity: 0.8;
+            min-width: 100px;
+            text-align: center;
+        }}
+        .page-nav .back-link {{
+            color: #44aaff;
+            text-decoration: none;
+            font-size: 12px;
+            margin-left: 10px;
+        }}
+        {dzi_note}
+    </style>
+</head>
+<body>
+    <div class="layout">
+        <div class="viewer-area">
+            <div class="controls">
+                <button id="btn-all" class="active" onclick="filterIssues('all')">Все</button>
+                <button id="btn-error" onclick="filterIssues('error')">Ошибки</button>
+                <button id="btn-warning" onclick="filterIssues('warning')">Предупр.</button>
+                <button id="btn-info" onclick="filterIssues('info')">Инфо</button>
+                <button onclick="resetView()">Сброс</button>
+            </div>
+            <div id="openseadragon"></div>
+            
+            <!-- Page Navigation -->
+            <div class="page-nav" id="pageNav">
+                <button id="btnPrev" onclick="goPage(-1)">◀ Предыдущая</button>
+                <span class="page-info">Страница <span id="pageNum">{target_page}</span> / <span id="totalPages">{total_pages or '?'}</span></span>
+                <select id="pageSelect" onchange="jumpPage(this.value)">
+                    {''.join([f'<option value="{i}"{" selected" if i == target_page else ""}>Стр. {i}</option>' for i in range(1, (total_pages or target_page or 1) + 1)])}
+                </select>
+                <button id="btnNext" onclick="goPage(1)">Следующая ▶</button>
+                <a href="{api_base or '.'}" class="back-link">← Dashboard</a>
+            </div>
+        </div>
+        
+        <div class="sidebar">
+            <div class="sidebar-header">
+                <h2>Замечания — Стр. {target_page}</h2>
+                <div class="stats">
+                    <div class="stat error">
+                        <div class="stat-value">{len([i for i in issues if i['severity']=='error'])}</div>
+                        <div class="stat-label">Ошибки</div>
+                    </div>
+                    <div class="stat warning">
+                        <div class="stat-value">{len([i for i in issues if i['severity']=='warning'])}</div>
+                        <div class="stat-label">Предупр.</div>
+                    </div>
+                    <div class="stat info">
+                        <div class="stat-value">{len([i for i in issues if i['severity']=='info'])}</div>
+                        <div class="stat-label">Инфо</div>
+                    </div>
+                </div>
+                <div class="legend">
+                    <div class="legend-item">
+                        <div class="legend-dot" style="background:#ff4444;"></div>
+                        <span>Ошибка</span>
+                    </div>
+                    <div class="legend-item">
+                        <div class="legend-dot" style="background:#ffaa00;"></div>
+                        <span>Предупреждение</span>
+                    </div>
+                    <div class="legend-item">
+                        <div class="legend-dot" style="background:#44aaff;"></div>
+                        <span>Информация</span>
+                    </div>
+                </div>
+            </div>
+            
+            <div class="issue-list" id="issueList">
+'''
+    
+    issue_counter = 0
+    for issue in issues:
+        issue_counter += 1
+        css_var = f"--card-color: {issue['color']}"
+        db_id_attr = f'data-db-id="{issue_db_ids[issue_counter-1]}"' if issue_db_ids and issue_counter <= len(issue_db_ids) else ''
+        has_api = 'true' if issue_db_ids else 'false'
+        html += f'''
+                <div class="issue-card" style="{css_var}" 
+                     data-severity="{issue['severity']}" data-id="{issue_counter}" {db_id_attr} data-has-api="{has_api}"
+                     onclick="focusIssue({issue_counter})">
+                    <div class="issue-type">
+                        <span class="issue-num">#{issue_counter}</span>
+                        {issue["type"]}
+                    </div>
+                    <div class="issue-message">{issue["message"][:250]}</div>
+                    <div class="issue-severity">{issue["severity"]}</div>
+                    <div class="feedback-bar" onclick="event.stopPropagation();">
+                        <button class="tp" onclick="submitFeedback(this, true)" title="Реальная проблема">✅ Да</button>
+                        <button class="fp" onclick="submitFeedback(this, false)" title="Ложное срабатывание">❌ Нет</button>
+                        <button class="ns" onclick="submitFeedback(this, null)" title="Не уверен">🤷 Не знаю</button>
+                    </div>
+                    <div class="feedback-status" id="fbstatus-{issue_counter if not issue_db_ids else issue_db_ids[issue_counter-1]}"></div>
+                </div>
+'''
+    
+    html += f'''
+            </div>
+        </div>
+    </div>
+    
+    <div id="tooltip"></div>
+    <div id="toast"></div>
+    
+    <script>
+        const OVERLAYS = {overlays_json};
+        let viewer;
+        let activeFilter = 'all';
+        
+        // Initialize OpenSeadragon
+        viewer = OpenSeadragon({{
+            id: "openseadragon",
+            prefixUrl: "https://cdn.jsdelivr.net/npm/openseadragon@4.1/build/openseadragon/images/",
+            tileSources: {tile_source},
+            showNavigationControl: true,
+            navigationControlAnchor: OpenSeadragon.ControlAnchor.TOP_RIGHT,
+            showZoomControl: true,
+            showHomeControl: true,
+            showFullPageControl: true,
+            zoomInButton: "zoom-in",
+            zoomOutButton: "zoom-out",
+            homeButton: "home",
+            fullPageButton: "full-page",
+            maxZoomPixelRatio: 10,
+            minZoomLevel: 0.1,
+            visibilityRatio: 0.5,
+            constrainDuringPan: true,
+        }});
+        
+        // Add overlays after viewer opens
+        viewer.addHandler('open', function() {{
+            addOverlays();
+        }});
+        
+        function addOverlays() {{
+            OVERLAYS.forEach((ov, idx) => {{
+                const el = document.createElement('div');
+                el.className = 'overlay-rect';
+                el.id = 'overlay-' + (idx + 1);
+                el.style.setProperty('--oc-color', ov.color);
+                el.setAttribute('data-severity', ov.severity);
+                
+                const num = document.createElement('div');
+                num.className = 'overlay-number';
+                num.textContent = '⚠ ' + ov.num;
+                el.appendChild(num);
+                
+                // Tooltip events
+                el.addEventListener('mouseenter', (e) => showTooltip(e, ov));
+                el.addEventListener('mousemove', moveTooltip);
+                el.addEventListener('mouseleave', hideTooltip);
+                el.addEventListener('click', () => focusIssue(ov.num));
+                
+                viewer.addOverlay({{
+                    element: el,
+                    location: new OpenSeadragon.Rect(ov.x, ov.y, ov.w, ov.h)
+                }});
+            }});
+        }}
+        
+        const tooltip = document.getElementById('tooltip');
+        
+        function showTooltip(e, ov) {{
+            tooltip.innerHTML = '<strong>#' + ov.num + ' — ' + ov.type + '</strong><br><span style="opacity:0.7">' + ov.severity.toUpperCase() + '</span><br><br>' + ov.message;
+            tooltip.style.display = 'block';
+            tooltip.style.borderColor = ov.color;
+        }}
+        
+        function moveTooltip(e) {{
+            tooltip.style.left = (e.clientX + 15) + 'px';
+            tooltip.style.top = (e.clientY + 15) + 'px';
+        }}
+        
+        function hideTooltip() {{
+            tooltip.style.display = 'none';
+        }}
+        
+        function filterIssues(severity) {{
+            activeFilter = severity;
+            
+            // Update buttons
+            document.querySelectorAll('.controls button').forEach(btn => btn.classList.remove('active'));
+            document.getElementById('btn-' + severity).classList.add('active');
+            
+            // Filter overlays
+            document.querySelectorAll('.overlay-rect').forEach(el => {{
+                const sev = el.getAttribute('data-severity');
+                el.style.display = (severity === 'all' || sev === severity) ? 'block' : 'none';
+            }});
+            
+            // Filter cards
+            document.querySelectorAll('.issue-card').forEach(card => {{
+                const sev = card.getAttribute('data-severity');
+                card.style.display = (severity === 'all' || sev === severity) ? 'block' : 'none';
+            }});
+        }}
+        
+        function focusIssue(num) {{
+            console.log('focusIssue called for num:', num);
+            selectIssue(num);
+            
+            // Find overlay and zoom to it
+            const overlay = OVERLAYS[num - 1];
+            console.log('Found overlay:', overlay);
+            if (!overlay) {{
+                console.warn('No overlay found for num', num);
+                return;
+            }}
+            
+            // Wait for viewer to be ready
+            if (!viewer || !viewer.viewport) {{
+                console.warn('Viewer not ready yet, retrying...');
+                setTimeout(() => focusIssue(num), 500);
+                return;
+            }}
+            
+            console.log('Viewer ready, fitting bounds to:', overlay);
+            const rect = new OpenSeadragon.Rect(
+                Math.max(0, overlay.x - 0.05), 
+                Math.max(0, overlay.y - 0.05),
+                Math.min(1 - overlay.x + 0.05, overlay.w + 0.1), 
+                Math.min(1 - overlay.y + 0.05, overlay.h + 0.1)
+            );
+            console.log('Rect:', rect);
+            viewer.viewport.fitBounds(rect);
+            console.log('fitBounds called');
+        }}
+        
+        function selectIssue(num) {{
+            // Highlight card
+            document.querySelectorAll('.issue-card').forEach(c => c.classList.remove('active'));
+            const card = document.querySelector(`.issue-card[data-id="${{num}}"]`);
+            if (card) {{
+                card.classList.add('active');
+                card.scrollIntoView({{behavior: 'smooth', block: 'center'}});
+            }}
+            
+            // Highlight overlay
+            document.querySelectorAll('.overlay-rect').forEach(el => {{
+                el.style.opacity = el.id === 'overlay-' + num ? '0.6' : '0.15';
+                el.style.borderWidth = el.id === 'overlay-' + num ? '4px' : '2px';
+            }});
+        }}
+        
+        function resetView() {{
+            viewer.viewport.goHome();
+            document.querySelectorAll('.issue-card').forEach(c => c.classList.remove('active'));
+            document.querySelectorAll('.overlay-rect').forEach(el => {{
+                el.style.opacity = '0.15';
+                el.style.borderWidth = '2px';
+            }});
+        }}
+        
+        // ===== FEEDBACK SYSTEM =====
+        async function submitFeedback(btn, isTP) {{
+            const card = btn.closest('.issue-card');
+            const hasApi = card.getAttribute('data-has-api') === 'true';
+            const dbId = card.getAttribute('data-db-id');
+            
+            if (!hasApi || !dbId) {{
+                showToast('API недоступен. Сохраните viewer через backend.', 'warning');
+                // Still update UI for demo
+                updateFeedbackUI(card, isTP, false);
+                return;
+            }}
+            
+            btn.disabled = true;
+            const originalText = btn.textContent;
+            btn.textContent = '...';
+            
+            try {{
+                const res = await fetch('/api/feedback', {{
+                    method: 'POST',
+                    headers: {{'Content-Type': 'application/json'}},
+                    body: JSON.stringify({{
+                        issue_id: parseInt(dbId),
+                        is_true_positive: isTP,
+                        action_taken: isTP === true ? 'fixed' : (isTP === false ? 'ignored' : 'not_sure')
+                    }})
+                }});
+                
+                if (res.ok) {{
+                    updateFeedbackUI(card, isTP, true);
+                    showToast('Feedback сохранён!', 'success');
+                }} else {{
+                    const err = await res.text();
+                    showToast('Ошибка: ' + err, 'error');
+                    btn.disabled = false;
+                    btn.textContent = originalText;
+                }}
+            }} catch(e) {{
+                showToast('Сетевой сбой: ' + e.message, 'error');
+                btn.disabled = false;
+                btn.textContent = originalText;
+            }}
+        }}
+        
+        function updateFeedbackUI(card, isTP, saved) {{
+            const statusDiv = card.querySelector('.feedback-status');
+            const btns = card.querySelector('.feedback-bar');
+            
+            // Remove all feedback classes
+            card.classList.remove('feedback-tp', 'feedback-fp');
+            
+            if (isTP === true) {{
+                card.classList.add('feedback-tp');
+                btns.innerHTML = '<span style="color:#00c864; font-size:12px;">✅ Подтверждено как реальная проблема</span>';
+                if (statusDiv) statusDiv.innerHTML = '<span class="tp">Сохранено' + (saved ? '' : ' (локально)') + '</span>';
+            }} else if (isTP === false) {{
+                card.classList.add('feedback-fp');
+                btns.innerHTML = '<span style="color:#ff4444; font-size:12px;">❌ Отклонено (ложное срабатывание)</span>';
+                if (statusDiv) statusDiv.innerHTML = '<span class="fp">Сохранено' + (saved ? '' : ' (локально)') + '</span>';
+            }} else {{
+                btns.innerHTML = '<span style="color:#888; font-size:12px;">🤷 Не уверен</span>';
+                if (statusDiv) statusDiv.innerHTML = '<span class="ns">Сохранено' + (saved ? '' : ' (локально)') + '</span>';
+            }}
+        }}
+        
+        function showToast(msg, type) {{
+            const toast = document.getElementById('toast');
+            toast.textContent = msg;
+            toast.style.borderColor = type === 'success' ? '#00c864' : (type === 'error' ? '#ff4444' : '#ffaa00');
+            toast.style.display = 'block';
+            setTimeout(() => {{ toast.style.display = 'none'; }}, 3000);
+        }}
+        
+        // ===== PAGE NAVIGATION =====
+        const PROJECT_ID = null;
+        const CURRENT_PAGE = null;
+        const TOTAL_PAGES = null;
+        const API_BASE = '';
+        
+        function updatePageNav() {{
+            const btnPrev = document.getElementById('btnPrev');
+            const btnNext = document.getElementById('btnNext');
+            const pageSelect = document.getElementById('pageSelect');
+            
+            if (btnPrev) btnPrev.disabled = CURRENT_PAGE <= 1;
+            if (btnNext) btnNext.disabled = TOTAL_PAGES && CURRENT_PAGE >= TOTAL_PAGES;
+            if (pageSelect) pageSelect.value = CURRENT_PAGE;
+            
+            // If no backend project ID, disable nav and show message
+            if (!PROJECT_ID) {{
+                if (btnPrev) btnPrev.disabled = true;
+                if (btnNext) btnNext.disabled = true;
+                const nav = document.getElementById('pageNav');
+                if (nav) nav.title = 'Page navigation requires backend server';
+            }}
+        }}
+        
+        function goPage(delta) {{
+            if (!PROJECT_ID || !TOTAL_PAGES) return;
+            const newPage = CURRENT_PAGE + delta;
+            if (newPage < 1 || newPage > TOTAL_PAGES) return;
+            jumpPage(newPage);
+        }}
+        
+        function jumpPage(pageNum) {{
+            if (!PROJECT_ID) {{
+                showToast('Navigation requires backend server', 'warning');
+                return;
+            }}
+            const base = API_BASE || '';
+            window.location.href = base + '/viewer/' + PROJECT_ID + '/' + pageNum;
+        }}
+        
+        // Initialize
+        updatePageNav();
+    </script>
+</body>
+</html>'''
+    
+    out_dir = folder / "web_viewer"
+    out_dir.mkdir(exist_ok=True)
+    
+    # Copy DZI tiles into web_viewer folder so relative paths work
+    if use_dzi:
+        src_files = folder / f"page_{target_page:03d}_files"
+        dst_files = out_dir / f"page_{target_page:03d}_files"
+        if src_files.exists():
+            if dst_files.exists():
+                shutil.rmtree(dst_files)
+            shutil.copytree(src_files, dst_files)
+            print(f"[INFO] Скопировано тайлов: {dst_files}")
+        else:
+            print(f"[WARN] Тайлы не найдены: {src_files}")
+            use_dzi = False
+    
+    out_path = out_dir / "index.html"
+    out_path.write_text(html, encoding="utf-8")
+    
+    print(f"[OK] Viewer создан: {out_path}")
+    print(f"      {'DZI tiles' if use_dzi else 'Direct PNG'}")
+    print(f"      Откройте в браузере: file://{out_path}")
+
+
+def main():
+    folder = Path(sys.argv[1]) if len(sys.argv) > 1 else Path("output_123")
+    page = int(sys.argv[2]) if len(sys.argv) > 2 else None
+    use_dzi = "--dzi" in sys.argv or "--no-dzi" not in sys.argv
+    generate_html(folder, page, use_dzi)
+
+
+if __name__ == "__main__":
+    main()

gost_dimension_validator.py

@@ -0,0 +1,235 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Валидатор ГОСТ-ов и размеров на чертежах.
+
+Проверяет OCR-результаты на:
+1. Найденные ГОСТ/СНиП/СП/ТУ — сверка с базой устаревших
+2. Размеры — валидация по типовым модулям и суммам
+3. Низкий confidence OCR — флаги для ручной проверки
+
+Использование:
+    python gost_dimension_validator.py <output_folder>
+"""
+
+import sys
+import json
+import re
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+# ------------------------------------------------------------------
+# База устаревших ГОСТов (пример — расширяется)
+# ------------------------------------------------------------------
+GOST_DATABASE = {
+    # Устаревшие ГОСТы → замена
+    "ГОСТ 21.101-97": {"status": "active", "name": "Система проектной документации"},
+    "ГОСТ 21.501-93": {"status": "obsolete", "replacement": "ГОСТ Р 21.1017-2022", "name": "Правила выполнения архитектурных чертежей"},
+    "ГОСТ 2.301-68": {"status": "active", "name": "Форматы"},
+    "ГОСТ 2.302-68": {"status": "obsolete", "replacement": "ГОСТ 2.302-2019", "name": "Масштабы"},
+    "ГОСТ 2.303-68": {"status": "obsolete", "replacement": "ГОСТ 2.303-2020", "name": "Линии"},
+    "ГОСТ 2.304-81": {"status": "obsolete", "replacement": "ГОСТ 2.304-2021", "name": "Шрифты чертежные"},
+    "ГОСТ 2.305-2008": {"status": "active", "name": "Изображения виды"},
+    "ГОСТ 2.307-2011": {"status": "active", "name": "Нанесение размеров"},
+    "СНиП II-22-81": {"status": "obsolete", "replacement": "СП 70.13330.2012", "name": "Каменные и армокаменные конструкции"},
+    "СНиП 2.01.07-85": {"status": "obsolete", "replacement": "СП 20.13330.2016", "name": "Нагрузки и воздействия"},
+    "СНиП 31-01-2003": {"status": "obsolete", "replacement": "СП 54.13330.2016", "name": "Жилые многоквартирные дома"},
+}
+
+# Типовые строительные модули (мм)
+CONSTRUCTION_MODULES = [100, 200, 300, 400, 500, 600, 1000, 1200, 1500, 1800, 2400, 3000, 3600, 4200, 5400, 6000, 6600]
+
+# ------------------------------------------------------------------
+# Парсеры
+# ------------------------------------------------------------------
+def extract_gosts(text: str) -> List[Tuple[str, int]]:
+    """Извлекает ГОСТ/СНиП/СП/ТУ из текста с позициями."""
+    patterns = [
+        r'ГОСТ\s*Р?\s*\d{1,5}(?:[-.]\d+)*(?:-\d{2,4})?',  # ГОСТ 12345-67, ГОСТ Р 21.1017-2022
+        r'СНиП\s*(?:[IVX]+[-.])?\s*\d{1,3}[-.]\d{1,3}[-.]?\d{0,4}',  # СНиП II-22-81, СНиП 31-01-2003
+        r'СП\s*\d{1,3}\.\d{1,6}\.\d{4}',  # СП 54.13330.2016
+        r'ТУ\s*\d{1,4}(?:[-/]\d+)*[-.]\d{4}',  # ТУ 400-... 
+    ]
+    found = []
+    for pat in patterns:
+        for m in re.finditer(pat, text, re.I):
+            found.append((m.group(0), m.start()))
+    return found
+
+
+def extract_dimensions(text: str) -> List[Tuple[str, float]]:
+    """Извлекает размеры в мм/м/см."""
+    found = []
+    # Основные размеры в мм (3600, 5400, 125.30)
+    for m in re.finditer(r'\b(\d{1,5}(?:[.,]\d{1,2})?)\s*м?[мм]?\b', text):
+        val = m.group(1).replace(',', '.')
+        try:
+            num = float(val)
+            if 10 <= num <= 50000:  # реалистичные строительные размеры
+                found.append((m.group(0), num))
+        except ValueError:
+            pass
+    return found
+
+
+def is_typical_module(dim: float, tolerance: float = 5.0) -> bool:
+    """Проверяет, кратен ли размер типовому модулю."""
+    for mod in CONSTRUCTION_MODULES:
+        if abs(dim - mod) < tolerance or abs(dim % mod) < tolerance:
+            return True
+    return False
+
+
+def validate_gost(gost: str) -> dict:
+    """Проверяет статус ГОСТа в базе."""
+    gost_norm = gost.strip().upper()
+    # Нормализация
+    gost_norm = re.sub(r'\s+', ' ', gost_norm)
+    
+    # Точное совпадение
+    if gost_norm in GOST_DATABASE:
+        info = GOST_DATABASE[gost_norm].copy()
+        info["gost"] = gost
+        return info
+    
+    # Нечёткий поиск (без года)
+    base = re.sub(r'-\d{2,4}$', '', gost_norm)
+    for key, info in GOST_DATABASE.items():
+        key_base = re.sub(r'-\d{2,4}$', '', key)
+        if base == key_base:
+            result = info.copy()
+            result["gost"] = gost
+            result["note"] = f"Найден по базовому номеру ({key})"
+            return result
+    
+    return {"gost": gost, "status": "unknown", "note": "Не найден в базе"}
+
+
+# ------------------------------------------------------------------
+# Основная логика
+# ------------------------------------------------------------------
+def validate_folder(folder: Path):
+    """Проверяет OCR-данные из full_ocr_results.json."""
+    ocr_path = folder / "full_ocr_results.json"
+    if not ocr_path.exists():
+        print(f"[ERR] Не найден {ocr_path}")
+        sys.exit(1)
+    
+    data = json.loads(ocr_path.read_text(encoding="utf-8"))
+    pages = data["pages"]
+    
+    print(f"[INFO] Проверка {len(pages)} страниц...\n")
+    
+    all_gosts = []
+    all_dims = []
+    low_confidence_items = []
+    
+    for page in pages:
+        page_num = page["page_number"]
+        
+        # --- 1. Проверка ГОСТ-ов ---
+        full_text = page.get("pdf_text_layer", "")
+        for line in page.get("ocr_lines", []):
+            full_text += " " + line["text"]
+        
+        gosts = extract_gosts(full_text)
+        for gost, pos in gosts:
+            info = validate_gost(gost)
+            all_gosts.append({
+                "page": page_num,
+                "gost": gost,
+                **info
+            })
+        
+        # --- 2. Проверка размеров ---
+        dims = extract_dimensions(full_text)
+        for dim_text, dim_val in dims:
+            is_typical = is_typical_module(dim_val)
+            all_dims.append({
+                "page": page_num,
+                "text": dim_text,
+                "value": dim_val,
+                "typical": is_typical,
+            })
+        
+        # --- 3. Низкий confidence OCR ---
+        for line in page.get("ocr_lines", []):
+            conf = line.get("confidence", 0)
+            if conf < 0.6:
+                low_confidence_items.append({
+                    "page": page_num,
+                    "text": line["text"],
+                    "confidence": conf,
+                    "bbox": line.get("bbox", []),
+                })
+    
+    # --- Вывод результатов ---
+    print("=" * 60)
+    print("ГОСТ/СНиП/СП/ТУ:")
+    print("=" * 60)
+    obsolete = [g for g in all_gosts if g["status"] == "obsolete"]
+    active = [g for g in all_gosts if g["status"] == "active"]
+    unknown = [g for g in all_gosts if g["status"] == "unknown"]
+    
+    if obsolete:
+        print(f"\n⚠️ УСТАРЕВШИЕ ({len(obsolete)}):")
+        for g in obsolete:
+            print(f"  Стр.{g['page']}: {g['gost']}")
+            print(f"    → Замена: {g.get('replacement', 'не указана')}")
+    if active:
+        print(f"\n✅ АКТУАЛЬНЫЕ ({len(active)}):")
+        for g in active[:10]:
+            print(f"  Стр.{g['page']}: {g['gost']} ({g.get('name', '')})")
+        if len(active) > 10:
+            print(f"  ... и ещё {len(active) - 10}")
+    if unknown:
+        print(f"\n❓ НЕИЗВЕСТНЫЕ ({len(unknown)}):")
+        for g in unknown[:5]:
+            print(f"  Стр.{g['page']}: {g['gost']}")
+    
+    print("\n" + "=" * 60)
+    print("РАЗМЕРЫ:")
+    print("=" * 60)
+    typical = [d for d in all_dims if d["typical"]]
+    atypical = [d for d in all_dims if not d["typical"]]
+    
+    print(f"\n✅ Типовые модули ({len(typical)}):")
+    for d in typical[:10]:
+        print(f"  Стр.{d['page']}: {d['text']} → {d['value']} мм")
+    
+    if atypical:
+        print(f"\n⚠️ НЕТИПОВЫЕ/ПРОВЕРИТЬ ({len(atypical)}):")
+        for d in atypical[:10]:
+            print(f"  Стр.{d['page']}: {d['text']} → {d['value']} мм (не кратен модулю)")
+    
+    print("\n" + "=" * 60)
+    print("НИЗКИЙ CONFIDENCE OCR (< 0.6):")
+    print("=" * 60)
+    if low_confidence_items:
+        print(f"\n⚠️ Найдено {len(low_confidence_items)} элементов для проверки:")
+        for item in low_confidence_items[:15]:
+            print(f"  Стр.{item['page']}: '{item['text']}' (conf={item['confidence']:.2f})")
+        if len(low_confidence_items) > 15:
+            print(f"  ... и ещё {len(low_confidence_items) - 15}")
+    else:
+        print("\n✅ Все элементы с высоким confidence")
+    
+    # --- Сохранение JSON ---
+    report = {
+        "gosts": {"obsolete": obsolete, "active": active, "unknown": unknown},
+        "dimensions": {"typical": typical, "atypical": atypical},
+        "low_confidence": low_confidence_items,
+    }
+    out_path = folder / "validation_report.json"
+    with open(out_path, "w", encoding="utf-8") as f:
+        json.dump(report, f, ensure_ascii=False, indent=2)
+    print(f"\n[INFO] Отчёт сохранён: {out_path}")
+
+
+def main():
+    folder = Path(sys.argv[1]) if len(sys.argv) > 1 else Path("output_123")
+    validate_folder(folder)
+
+
+if __name__ == "__main__":
+    main()

layout_detector.py

@@ -0,0 +1,298 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Layout Detector — разделение страницы чертежа на зоны.
+
+Зоны:
+- "drawing" — схемы, виды, разрезы (линии + текст, разрежено)
+- "table" — таблицы (плотные линии в сетке)
+- "title_block" — штамп (нижний правый угол или низ страницы)
+- "notes" — примечания, текстовые блоки
+- "legend" — легенда/условные обозначения
+
+Алгоритм:
+1. Находит все линии на странице
+2. Находит прямоугольники = таблицы
+3. Анализирует плотность OCR текста
+4. Классифицирует регионы
+"""
+
+import sys
+import json
+from pathlib import Path
+from typing import List, Dict, Tuple
+import cv2
+import numpy as np
+from PIL import Image
+
+
+def find_all_lines(img_gray: np.ndarray, min_length: int = 40):
+    """Находит все прямые линии (горизонтальные и вертикальные)."""
+    _, binary = cv2.threshold(img_gray, 180, 255, cv2.THRESH_BINARY_INV)
+    h, w = binary.shape
+    lines = []
+    
+    # Горизонтальные
+    for y in range(h):
+        row = binary[y, :]
+        in_line = False
+        start = 0
+        for x in range(w):
+            if row[x] > 128:
+                if not in_line:
+                    in_line = True
+                    start = x
+            else:
+                if in_line:
+                    length = x - start
+                    if length >= min_length:
+                        lines.append(("h", start, y, x-1, y))
+                    in_line = False
+        if in_line:
+            length = w - start
+            if length >= min_length:
+                lines.append(("h", start, y, w-1, y))
+    
+    # Вертикальные
+    for x in range(w):
+        col = binary[:, x]
+        in_line = False
+        start = 0
+        for y in range(h):
+            if col[y] > 128:
+                if not in_line:
+                    in_line = True
+                    start = y
+            else:
+                if in_line:
+                    length = y - start
+                    if length >= min_length:
+                        lines.append(("v", x, start, x, y-1))
+                    in_line = False
+        if in_line:
+            length = h - start
+            if length >= min_length:
+                lines.append(("v", x, start, x, h-1))
+    
+    return lines
+
+
+def find_rectangles(lines: List[Tuple], min_size: int = 100) -> List[Dict]:
+    """Находит прямоугольники, образованные пересечением линий."""
+    horiz = [(l[1], l[2], l[3], l[4]) for l in lines if l[0] == "h"]
+    vert = [(l[1], l[2], l[3], l[4]) for l in lines if l[0] == "v"]
+    
+    # Группируем горизонтальные по Y
+    from collections import defaultdict
+    h_by_y = defaultdict(list)
+    for x1, y1, x2, y2 in horiz:
+        h_by_y[y1].append((x1, x2))
+    
+    # Группируем вертикальные по X
+    v_by_x = defaultdict(list)
+    for x1, y1, x2, y2 in vert:
+        v_by_x[x1].append((y1, y2))
+    
+    rects = []
+    # Ищем пары горизонтальных линий с общими вертикальными
+    y_vals = sorted(h_by_y.keys())
+    for i in range(len(y_vals)):
+        for j in range(i+1, len(y_vals)):
+            y_top = y_vals[i]
+            y_bottom = y_vals[j]
+            # Ищем общий X-интервал
+            for x1_a, x2_a in h_by_y[y_top]:
+                for x1_b, x2_b in h_by_y[y_bottom]:
+                    x_left = max(x1_a, x1_b)
+                    x_right = min(x2_a, x2_b)
+                    if x_right - x_left < min_size:
+                        continue
+                    # Проверяем, есть ли вертикальные линии на x_left и x_right
+                    has_left = any(y_top <= y_bottom and not (y2 < y_top or y1 > y_bottom) 
+                                   for y1, y2 in v_by_x.get(x_left, []))
+                    has_right = any(y_top <= y_bottom and not (y2 < y_top or y1 > y_bottom)
+                                    for y1, y2 in v_by_x.get(x_right, []))
+                    if has_left and has_right:
+                        rects.append({
+                            "x": x_left, "y": y_top,
+                            "w": x_right - x_left, "h": y_bottom - y_top
+                        })
+    
+    # Фильтруем вложенные прямоугольники (оставляем только внешние)
+    filtered = []
+    for r in rects:
+        is_inner = False
+        for other in rects:
+            if r is other:
+                continue
+            if (r["x"] > other["x"] and r["y"] > other["y"] and
+                r["x"] + r["w"] < other["x"] + other["w"] and
+                r["y"] + r["h"] < other["y"] + other["h"]):
+                is_inner = True
+                break
+        if not is_inner:
+            filtered.append(r)
+    
+    return filtered
+
+
+def classify_regions(rects: List[Dict], ocr_lines: List[Dict], img_w: int, img_h: int) -> List[Dict]:
+    """Классифицирует регионы страницы."""
+    regions = []
+    
+    # 1. Таблицы = большие прямоугольники с высокой плотностью линий
+    for r in rects:
+        area = r["w"] * r["h"]
+        # Считаем OCR строки внутри
+        texts_in = [t for t in ocr_lines
+                    if r["x"] <= t["cx"] <= r["x"] + r["w"]
+                    and r["y"] <= t["cy"] <= r["y"] + r["h"]]
+        density = len(texts_in) / (area / 1000000)  # текстов на мегапиксель
+        
+        if density > 20:  # высокая плотность = таблица
+            regions.append({
+                "type": "table",
+                "bbox": [r["x"], r["y"], r["x"]+r["w"], r["y"]+r["h"]],
+                "density": density,
+                "text_count": len(texts_in)
+            })
+    
+    # 2. Определяем чертежи = области с линиями и текстом, но без плотной сетки
+    # Для простоты: левая половина, не покрытая таблицами
+    # Найдём ограничивающий bbox для всех "чертёжных" текстов
+    drawing_texts = [t for t in ocr_lines if t["cy"] < img_h * 0.75 and t["cx"] < img_w * 0.6]
+    if drawing_texts:
+        xs = [t["cx"] for t in drawing_texts]
+        ys = [t["cy"] for t in drawing_texts]
+        # Расширяем на 200px
+        dx = [t["cx"] - t["x1"] for t in drawing_texts if "x1" in t]
+        max_w = max(dx) if dx else 100
+        regions.append({
+            "type": "drawing",
+            "bbox": [max(0, min(xs)-max_w), max(0, min(ys)-100),
+                     min(img_w, max(xs)+max_w), min(img_h, max(ys)+100)],
+            "text_count": len(drawing_texts)
+        })
+    
+    # 3. Штамп = низ страницы, мелкий текст
+    title_texts = [t for t in ocr_lines if t["cy"] > img_h * 0.85]
+    if title_texts:
+        xs = [t["cx"] for t in title_texts]
+        ys = [t["cy"] for t in title_texts]
+        regions.append({
+            "type": "title_block",
+            "bbox": [min(xs)-50, min(ys)-50, max(xs)+50, max(ys)+50],
+            "text_count": len(title_texts)
+        })
+    
+    # 4. Примечания = текстовые блоки
+    note_keywords = ["примечание", "общие указания", "границы", "размеры"]
+    note_texts = [t for t in ocr_lines
+                  if any(kw in t["text"].lower() for kw in note_keywords)]
+    if note_texts:
+        xs = [t["cx"] for t in note_texts]
+        ys = [t["cy"] for t in note_texts]
+        regions.append({
+            "type": "notes",
+            "bbox": [min(xs)-100, min(ys)-100, max(xs)+100, max(ys)+100],
+            "text_count": len(note_texts)
+        })
+    
+    return regions
+
+
+def detect_layout(png_path: Path, ocr_path: Path) -> Dict:
+    """Основная функция layout detection."""
+    img = cv2.imread(str(png_path), cv2.IMREAD_GRAYSCALE)
+    h, w = img.shape[:2]
+    
+    # Загрузить OCR
+    ocr = json.loads(ocr_path.read_text(encoding="utf-8"))
+    
+    # Собрать все OCR lines с координатами
+    all_texts = []
+    for page in ocr.get("pages", []):
+        for line in page.get("ocr_lines", []):
+            bbox = line.get("bbox", [])
+            if not bbox:
+                continue
+            if isinstance(bbox[0], list):
+                xs = [p[0] for p in bbox]
+                ys = [p[1] for p in bbox]
+            else:
+                xs = [bbox[0], bbox[2]]
+                ys = [bbox[1], bbox[3]]
+            all_texts.append({
+                "text": line["text"],
+                "cx": sum(xs)/len(xs),
+                "cy": sum(ys)/len(ys),
+                "x1": min(xs), "y1": min(ys),
+                "x2": max(xs), "y2": max(ys),
+                "bbox": bbox
+            })
+    
+    # Найти линии
+    lines = find_all_lines(img)
+    print(f"[INFO] Найдено {len(lines)} линий")
+    
+    # Найти прямоугольники
+    rects = find_rectangles(lines)
+    print(f"[INFO] Найдено {len(rects)} прямоугольников")
+    
+    # Классифицировать
+    regions = classify_regions(rects, all_texts, w, h)
+    print(f"[INFO] Классифицировано {len(regions)} регионов")
+    for r in regions:
+        print(f"  {r['type']}: bbox={r['bbox']}, texts={r.get('text_count', 0)}")
+    
+    return {
+        "image_size": [w, h],
+        "regions": regions,
+        "rectangles": rects,
+        "line_count": len(lines)
+    }
+
+
+def visualize_layout(png_path: Path, layout: Dict, out_path: Path):
+    """Рисует зоны на изображении."""
+    img = Image.open(png_path)
+    draw = ImageDraw.Draw(img)
+    colors = {
+        "table": "blue",
+        "drawing": "green",
+        "title_block": "purple",
+        "notes": "orange"
+    }
+    
+    for region in layout["regions"]:
+        x1, y1, x2, y2 = region["bbox"]
+        color = colors.get(region["type"], "red")
+        draw.rectangle([x1, y1, x2, y2], outline=color, width=4)
+        draw.text((x1+5, y1+5), region["type"], fill=color)
+    
+    img.save(out_path)
+    print(f"[OK] Layout visualization: {out_path}")
+
+
+def main():
+    if len(sys.argv) < 3:
+        print("Usage: python layout_detector.py <png> <ocr_json>")
+        sys.exit(1)
+    
+    png = Path(sys.argv[1])
+    ocr = Path(sys.argv[2])
+    out_json = png.parent / "layout.json"
+    out_png = png.parent / f"{png.stem}_layout.png"
+    
+    layout = detect_layout(png, ocr)
+    
+    with open(out_json, "w", encoding="utf-8") as f:
+        json.dump(layout, f, ensure_ascii=False, indent=2)
+    print(f"[OK] Layout JSON: {out_json}")
+    
+    visualize_layout(png, layout, out_png)
+
+
+if __name__ == "__main__":
+    from PIL import ImageDraw
+    main()

multi_element_extractor.py

@@ -0,0 +1,183 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Multi-Element Extractor — извлечение разных типов элементов из чертежа.
+
+Использует layout zones и OCR для извлечения:
+- dimensions: размеры (числа рядом с линиями в зоне drawing)
+- positions: позиции арматуры (П-1, X-1, etc.)
+- gosts: ссылки на ГОСТ
+- steel_grades: марки стали (A500C, B30, etc.)
+- elevations: отметки уровней (-1.060, etc.)
+- beam_labels: Балка Б-1, Б-2, Б-3
+- table_data: структурированные таблицы (позиция → длина, масса, etc.)
+"""
+
+import sys
+import json
+import re
+from pathlib import Path
+from typing import List, Dict
+
+
+def extract_from_zone(ocr_lines: List[Dict], zone_type: str, zone_bbox: List[int]) -> Dict:
+    """Извлекает элементы из конкретной зоны."""
+    results = {
+        "dimensions": [],
+        "positions": [],
+        "gosts": [],
+        "steel_grades": [],
+        "elevations": [],
+        "beam_labels": [],
+        "table_rows": []
+    }
+    
+    x1, y1, x2, y2 = zone_bbox
+    zone_texts = [t for t in ocr_lines
+                  if x1 <= t["cx"] <= x2 and y1 <= t["cy"] <= y2]
+    
+    for t in zone_texts:
+        txt = t["text"].strip()
+        
+        # ГОСТ
+        if re.search(r'ГОС\s*T?\s*\d+', txt):
+            results["gosts"].append({"text": txt, "bbox": t["bbox"]})
+        
+        # Марки стали
+        if re.search(r'A500C|B30|C\d+', txt, re.IGNORECASE):
+            results["steel_grades"].append({"text": txt, "bbox": t["bbox"]})
+        
+        # Балки
+        if re.match(r'Балка\s+Б-\d+', txt):
+            results["beam_labels"].append({"text": txt, "bbox": t["bbox"]})
+        
+        # Позиции (П-1, X-1, etc.)
+        if re.match(r'^[ПX]-\d+$', txt):
+            results["positions"].append({"text": txt, "bbox": t["bbox"]})
+        
+        # Отметки уровней
+        if re.match(r'^-?\d+[,.]\d+$', txt) and float(txt.replace(',', '.').replace('−', '-')) < 10:
+            results["elevations"].append({"text": txt, "bbox": t["bbox"]})
+        
+        # Размеры: только целые числа 2-4 цифры (исключаем мелкие фрагменты)
+        if zone_type == "drawing" and re.match(r'^\d{2,4}$', txt) and txt not in ('00', '000', '006'):
+            results["dimensions"].append({"text": txt, "bbox": t["bbox"]})
+    
+    # Для таблиц: структурируем
+    if zone_type == "table":
+        results["table_rows"] = structure_table(zone_texts)
+    
+    return results
+
+
+def structure_table(zone_texts: List[Dict]) -> List[Dict]:
+    """Простая структуризация таблицы: группировка по строкам (по Y)."""
+    if not zone_texts:
+        return []
+    
+    # Сортируем по Y
+    sorted_texts = sorted(zone_texts, key=lambda t: t["cy"])
+    
+    # Группируем по близости Y (±20px)
+    rows = []
+    current_row = []
+    last_y = None
+    for t in sorted_texts:
+        if last_y is None or abs(t["cy"] - last_y) < 20:
+            current_row.append(t)
+        else:
+            if current_row:
+                # Сортируем по X
+                current_row.sort(key=lambda x: x["cx"])
+                rows.append({"cells": [c["text"] for c in current_row]})
+            current_row = [t]
+        last_y = t["cy"]
+    
+    if current_row:
+        current_row.sort(key=lambda x: x["cx"])
+        rows.append({"cells": [c["text"] for c in current_row]})
+    
+    return rows
+
+
+def extract_all_elements(png_path: Path, ocr_path: Path, layout_path: Path) -> Dict:
+    """Извлекает все элементы по зонам."""
+    ocr = json.loads(ocr_path.read_text(encoding="utf-8"))
+    layout = json.loads(layout_path.read_text(encoding="utf-8"))
+    
+    # Собрать все OCR lines с координатами
+    all_texts = []
+    for page in ocr.get("pages", []):
+        for line in page.get("ocr_lines", []):
+            bbox = line.get("bbox", [])
+            if not bbox:
+                continue
+            if isinstance(bbox[0], list):
+                xs = [p[0] for p in bbox]
+                ys = [p[1] for p in bbox]
+            else:
+                xs = [bbox[0], bbox[2]]
+                ys = [bbox[1], bbox[3]]
+            all_texts.append({
+                "text": line["text"],
+                "cx": sum(xs)/len(xs),
+                "cy": sum(ys)/len(ys),
+                "bbox": bbox
+            })
+    
+    # Извлечь по зонам
+    all_results = {
+        "dimensions": [],
+        "positions": [],
+        "gosts": [],
+        "steel_grades": [],
+        "elevations": [],
+        "beam_labels": [],
+        "tables": []
+    }
+    
+    for region in layout.get("regions", []):
+        zone_results = extract_from_zone(all_texts, region["type"], region["bbox"])
+        for key in all_results:
+            if key in zone_results:
+                all_results[key].extend(zone_results[key])
+    
+    # Убрать дубликаты
+    for key in all_results:
+        seen = set()
+        unique = []
+        for item in all_results[key]:
+            if item["text"] not in seen:
+                seen.add(item["text"])
+                unique.append(item)
+        all_results[key] = unique
+    
+    return all_results
+
+
+def main():
+    if len(sys.argv) < 4:
+        print("Usage: python multi_element_extractor.py <png> <ocr_json> <layout_json>")
+        sys.exit(1)
+    
+    png = Path(sys.argv[1])
+    ocr = Path(sys.argv[2])
+    layout = Path(sys.argv[3])
+    
+    results = extract_all_elements(png, ocr, layout)
+    
+    out = png.parent / "elements.json"
+    with open(out, "w", encoding="utf-8") as f:
+        json.dump(results, f, ensure_ascii=False, indent=2)
+    
+    print(f"[OK] Elements saved: {out}")
+    for key, items in results.items():
+        print(f"  {key}: {len(items)} items")
+        for item in items[:5]:
+            print(f"    {item['text']}")
+        if len(items) > 5:
+            print(f"    ... and {len(items)-5} more")
+
+
+if __name__ == "__main__":
+    main()

ocr_qwen.py

@@ -0,0 +1,233 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+OCR через Alibaba Cloud qwen-vl-ocr API.
+
+Использование:
+    from ocr_qwen import run_ocr
+    results = run_ocr(image_path)
+
+Требует DASHSCOPE_API_KEY в .env
+"""
+
+import os
+import json
+import base64
+import io
+from pathlib import Path
+from typing import List, Dict, Tuple
+from PIL import Image
+from openai import OpenAI
+
+# Загрузить ключ
+_API_KEY = None
+_BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
+_MODEL = "qwen-vl-ocr"
+
+def _load_key():
+    global _API_KEY
+    if _API_KEY:
+        return _API_KEY
+    
+    # Попробовать .env
+    env_candidates = [
+        Path(__file__).parent / ".env",
+        Path(__file__).parent.parent / ".env",
+        Path(__file__).parent.parent.parent / ".env",
+    ]
+    for env_path in env_candidates:
+        if env_path.exists():
+            for line in env_path.read_text().splitlines():
+                if line.startswith("DASHSCOPE_API_KEY="):
+                    _API_KEY = line.split("=", 1)[1].strip()
+                    os.environ["DASHSCOPE_API_KEY"] = _API_KEY
+                    return _API_KEY
+    
+    _API_KEY = os.environ.get("DASHSCOPE_API_KEY")
+    return _API_KEY
+
+
+def resize_image(image_path: Path, max_size: int = 2048) -> Tuple[str, float, Tuple[int, int]]:
+    """
+    Уменьшает изображение до max_size по длинной стороне.
+    Возвращает: (base64_string, scale_factor, (orig_w, orig_h))
+    """
+    img = Image.open(image_path)
+    orig_w, orig_h = img.size
+    
+    # Если уже меньше — не менять
+    if max(orig_w, orig_h) <= max_size:
+        with open(image_path, "rb") as f:
+            b64 = base64.b64encode(f.read()).decode("utf-8")
+        return b64, 1.0, (orig_w, orig_h)
+    
+    # Вычислить новый размер
+    scale = max_size / max(orig_w, orig_h)
+    new_w = int(orig_w * scale)
+    new_h = int(orig_h * scale)
+    
+    img_resized = img.resize((new_w, new_h), Image.LANCZOS)
+    
+    # Сохранить в буфер
+    buf = io.BytesIO()
+    img_resized.save(buf, format="PNG")
+    b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
+    
+    return b64, scale, (orig_w, orig_h)
+
+
+def encode_image(image_path: Path) -> str:
+    with open(image_path, "rb") as f:
+        return base64.b64encode(f.read()).decode("utf-8")
+
+
+def parse_qwen_response(raw_text: str) -> List[Dict]:
+    """Парсит JSON из ответа qwen-vl-ocr."""
+    import re
+    text = raw_text.strip()
+    
+    # Удалить markdown code blocks ```json ... ```
+    if text.startswith("```"):
+        lines = text.splitlines()
+        start = 0
+        end = len(lines)
+        for i, line in enumerate(lines):
+            if line.strip().startswith("```") and start == 0:
+                start = i + 1
+            elif line.strip() == "```" and start > 0:
+                end = i
+                break
+        text = "\n".join(lines[start:end]).strip()
+    
+    # Робастный парсинг: извлекаем каждый объект отдельно через regex
+    results = []
+    # Шаблон: {"text": "...", "rotate_rect": [num, num, num, num, num]}
+    pattern = r'\{\s*"text":\s*"([^"]*)"\s*,\s*"rotate_rect":\s*\[\s*(-?\d+)\s*,\s*(-?\d+)\s*,\s*(-?\d+)\s*,\s*(-?\d+)\s*,\s*(-?\d+)\s*\]\s*\}'
+    
+    for match in re.finditer(pattern, text):
+        txt = match.group(1)
+        x, y, w, h, angle = int(match.group(2)), int(match.group(3)), int(match.group(4)), int(match.group(5)), int(match.group(6))
+        results.append({
+            "text": txt,
+            "rotate_rect": [x, y, w, h, angle]
+        })
+    
+    if not results:
+        # Fallback: попробовать стандартный JSON парсинг
+        try:
+            json_match = re.search(r'\[[\s\S]*\]', text)
+            if json_match:
+                data = json.loads(json_match.group(0))
+                if isinstance(data, list):
+                    return data
+        except Exception:
+            pass
+        print(f"[WARN] Regex parser не нашёл объекты, JSON тоже не распарсился")
+        print(f"[WARN] Text preview: {text[:200]}")
+    
+    return results
+
+
+def run_ocr(image_path: Path, verbose: bool = False) -> List[Dict]:
+    """
+    Запускает qwen-vl-ocr на изображении.
+    
+    Returns:
+        Список словарей: {
+            "text": str,
+            "bbox": [x1, y1, x2, y2, angle],  # rotate_rect format
+            "confidence": float  # estimated
+        }
+    """
+    api_key = _load_key()
+    if not api_key:
+        raise RuntimeError("DASHSCOPE_API_KEY not found in .env or environment")
+    
+    client = OpenAI(api_key=api_key, base_url=_BASE_URL)
+    
+    # Уменьшить изображение для экономии токенов
+    b64, scale, (orig_w, orig_h) = resize_image(image_path, max_size=2048)
+    data_url = f"data:image/png;base64,{b64}"
+    
+    if verbose:
+        orig_size = image_path.stat().st_size / 1024
+        print(f"[qwen-ocr] Отправка {image_path.name} (orig {orig_w}x{orig_h}, scale={scale:.2f}, {orig_size:.0f} KB)...", flush=True)
+    
+    response = client.chat.completions.create(
+        model=_MODEL,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": (
+                            "Распознай все текстовые элементы на этом чертеже. "
+                            "Для каждого текста верни ОТДЕЛЬНЫЙ JSON-объект с полями: text, rotate_rect [x,y,w,h,angle]. "
+                            "ВАЖНО: каждый текст — отдельный объект, без дублирующихся ключей в одном объекте. "
+                            "Пример правильного формата:\n"
+                            '[{"text": "Бетон", "rotate_rect": [100, 50, 30, 10, 0]}, {"text": "В30", "rotate_rect": [100, 65, 20, 10, 0]}]'
+                            "\nОтветь строго в формате JSON-массива без markdown."
+                        ),
+                    },
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                ],
+            }
+        ],
+        temperature=0.1,
+        max_tokens=8192,
+    )
+    
+    raw = response.choices[0].message.content.strip()
+    
+    # Сохранить raw для отладки
+    debug_path = image_path.parent / f"{image_path.stem}_qwen_raw.txt"
+    debug_path.write_text(raw, encoding="utf-8")
+    
+    items = parse_qwen_response(raw)
+    
+    # Конвертировать rotate_rect в наш формат, масштабируя обратно к оригиналу
+    results = []
+    for item in items:
+        rect = item.get("rotate_rect", [0, 0, 0, 0, 0])
+        if len(rect) >= 4:
+            x, y, w, h = rect[0], rect[1], rect[2], rect[3]
+            # Масштабировать обратно к оригинальному размеру
+            if scale != 1.0:
+                x = round(x / scale)
+                y = round(y / scale)
+                w = round(w / scale)
+                h = round(h / scale)
+            # bbox: [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]
+            bbox = [[x, y], [x + w, y], [x + w, y + h], [x, y + h]]
+        else:
+            bbox = None
+        
+        results.append({
+            "text": item.get("text", ""),
+            "bbox": bbox,
+            "confidence": 0.95,  # qwen-vl-ocr не возвращает confidence, ставим высокий
+            "source": "qwen-vl-ocr"
+        })
+    
+    if verbose:
+        print(f"[qwen-ocr] Найдено {len(results)} элементов")
+    
+    return results
+
+
+if __name__ == "__main__":
+    import sys
+    if len(sys.argv) < 2:
+        print("Usage: python ocr_qwen.py <image.png>")
+        sys.exit(1)
+    
+    image_path = Path(sys.argv[1])
+    results = run_ocr(image_path, verbose=True)
+    
+    print(f"\nНайдено {len(results)} текстовых элементов:")
+    for r in results[:20]:
+        print(f"  '{r['text']}' bbox={r['bbox']}")
+    
+    if len(results) > 20:
+        print(f"  ... и ещё {len(results) - 20}")

preprocess_for_ocr.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Предобработка PNG для улучшения OCR размерных чисел.
+
+Алгоритм:
+1. CLAHE — локальное повышение контраста
+2. Unsharp mask — повышение резкости
+3. Инвертирование (опционально для некоторых OCR)
+4. Масштабирование x2 (если исходное маленькое)
+"""
+
+import sys
+from pathlib import Path
+import cv2
+import numpy as np
+
+
+def preprocess_for_ocr(img_path: Path, out_path: Path, scale: float = 2.0):
+    img = cv2.imread(str(img_path), cv2.IMREAD_GRAYSCALE)
+    if img is None:
+        raise RuntimeError(f"Cannot load {img_path}")
+    
+    # Масштабирование
+    if scale != 1.0:
+        h, w = img.shape
+        img = cv2.resize(img, (int(w*scale), int(h*scale)), interpolation=cv2.INTER_CUBIC)
+    
+    # CLAHE (локальный контраст)
+    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
+    img = clahe.apply(img)
+    
+    # Unsharp mask
+    gaussian = cv2.GaussianBlur(img, (0,0), 3)
+    img = cv2.addWeighted(img, 1.5, gaussian, -0.5, 0)
+    
+    # Нормализация
+    img = cv2.normalize(img, None, 0, 255, cv2.NORM_MINMAX)
+    
+    cv2.imwrite(str(out_path), img)
+    print(f"[OK] Предобработка сохранена: {out_path}")
+
+
+if __name__ == "__main__":
+    if len(sys.argv) < 2:
+        print("Usage: python preprocess_for_ocr.py <png>")
+        sys.exit(1)
+    
+    png = Path(sys.argv[1])
+    out = png.parent / f"{png.stem}_preproc.png"
+    preprocess_for_ocr(png, out)

process_any_pdf.py

@@ -2,14 +2,21 @@
 # -*- coding: utf-8 -*-
 """
 Универсальное распознавание PDF в указанную папку.
+Поддерживает:
+  - RapidOCR (локально, быстро)
+  - RapidOCR + tiling (для больших чертежей)
+  - qwen-vl-ocr (API, точнее)
+
 Использование:
-    python process_any_pdf.py <pdf_file> <output_folder_name>
+    python process_any_pdf.py <pdf_file> <output_folder> [--use-qwen] [--use-tiling]
 """
 
 import sys
 import json
+import re
 import fitz
 from pathlib import Path
+from PIL import Image
 from rapidocr_onnxruntime import RapidOCR
 
 # ------------------------------------------------------------------
@@ -17,19 +24,95 @@ from rapidocr_onnxruntime import RapidOCR
 # ------------------------------------------------------------------
 DPI = 300
 BATCH_SIZE = 5
+TILE_SIZE = 2000
+TILE_OVERLAP = 200
 
 engine = RapidOCR()
 
+# qwen-vl-ocr lazy import
+try:
+    from ocr_qwen import run_ocr as qwen_ocr
+    QWEN_AVAILABLE = True
+except ImportError:
+    QWEN_AVAILABLE = False
+
+
 # ------------------------------------------------------------------
-def process_pdf(pdf_path: Path, out_dir: Path):
+# Tiling OCR helpers
+# ------------------------------------------------------------------
+def _make_tiles(img: Image.Image, tile_size: int = 2000, overlap: int = 200):
+    w, h = img.size
+    tiles = []
+    step = tile_size - overlap
+    for y in range(0, h, step):
+        for x in range(0, w, step):
+            x2 = min(x + tile_size, w)
+            y2 = min(y + tile_size, h)
+            tiles.append((x, y, img.crop((x, y, x2, y2))))
+    return tiles
+
+
+def _bbox_iou(a, b):
+    def _rect(box):
+        if isinstance(box[0], list):
+            xs = [p[0] for p in box]
+            ys = [p[1] for p in box]
+            return min(xs), min(ys), max(xs), max(ys)
+        return box[0], box[1], box[2], box[3]
+
+    ax1, ay1, ax2, ay2 = _rect(a)
+    bx1, by1, bx2, by2 = _rect(b)
+    ix1, iy1 = max(ax1, bx1), max(ay1, by1)
+    ix2, iy2 = min(ax2, bx2), min(ay2, by2)
+    if ix2 <= ix1 or iy2 <= iy1:
+        return 0.0
+    inter = (ix2 - ix1) * (iy2 - iy1)
+    union = (ax2 - ax1) * (ay2 - ay1) + (bx2 - bx1) * (by2 - by1) - inter
+    return inter / union if union > 0 else 0.0
+
+
+def run_tiling_ocr(img_path: Path, conf_threshold: float = 0.5):
+    """Запускает RapidOCR по кропам и объединяет результаты."""
+    img = Image.open(img_path)
+    tiles = _make_tiles(img, TILE_SIZE, TILE_OVERLAP)
+    all_results = []
+    for off_x, off_y, crop in tiles:
+        tmp = f"/tmp/tile_ocr.png"
+        crop.save(tmp)
+        res = engine(tmp)
+        if res and res[0]:
+            for item in res[0]:
+                box, txt, score = item
+                if score < conf_threshold:
+                    continue
+                shifted = [[pt[0] + off_x, pt[1] + off_y] for pt in box]
+                all_results.append({"text": txt, "confidence": float(score), "bbox": shifted})
+
+    # Дедупликация по IoU
+    unique = []
+    for r in sorted(all_results, key=lambda x: -x["confidence"]):
+        is_dup = any(_bbox_iou(r["bbox"], u["bbox"]) > 0.5 for u in unique)
+        if not is_dup:
+            unique.append(r)
+    return unique
+
+# ------------------------------------------------------------------
+def process_pdf(pdf_path: Path, out_dir: Path, use_qwen: bool = False, use_tiling: bool = False):
     out_dir.mkdir(parents=True, exist_ok=True)
     doc = fitz.open(pdf_path)
     total = len(doc)
-    print(f"=== PDF: {pdf_path.name} | Страниц: {total} -> {out_dir} ===\n")
+    print(f"=== PDF: {pdf_path.name} | Страниц: {total} -> {out_dir} ===")
+    if use_qwen:
+        print(f"[INFO] OCR engine: qwen-vl-ocr (API)")
+    elif use_tiling:
+        print(f"[INFO] OCR engine: RapidOCR + tiling ({TILE_SIZE}px tiles)")
+    else:
+        print(f"[INFO] OCR engine: RapidOCR (local)")
+    print()
     
     all_pages = []
     for i in range(total):
-        print(f"[{i+1}/{total}] Рендер + OCR ...", end=" ")
+        print(f"[{i+1}/{total}] Рендер + OCR ...", end=" ", flush=True)
         page = doc.load_page(i)
         raw_text = page.get_text("text").strip()
         
@@ -38,6 +121,40 @@ def process_pdf(pdf_path: Path, out_dir: Path):
         img_path = out_dir / f"page_{i+1:03d}.png"
         pix.save(img_path)
         
+        # Выбор OCR engine
+        if use_qwen and QWEN_AVAILABLE:
+            try:
+                ocr_lines = qwen_ocr(img_path, verbose=False)
+                print(f"qwen-ocr строк: {len(ocr_lines)}")
+            except Exception as e:
+                print(f"qwen-ocr ERR: {e}, fallback to RapidOCR")
+                ocr_lines = _run_rapidocr(img_path)
+                print(f"RapidOCR строк: {len(ocr_lines)}")
+        elif use_tiling:
+            ocr_lines = run_tiling_ocr(img_path)
+            print(f"Tiling OCR строк: {len(ocr_lines)}")
+        else:
+            ocr_lines = _run_rapidocr(img_path)
+            print(f"RapidOCR строк: {len(ocr_lines)}")
+        
+        all_pages.append({
+            "page_number": i + 1,
+            "image": str(img_path.name),
+            "pdf_text_layer": raw_text,
+            "ocr_lines": ocr_lines,
+            "ocr_line_count": len(ocr_lines)
+        })
+        
+        if (i + 1) % BATCH_SIZE == 0 or i == total - 1:
+            with open(out_dir / "full_ocr_results.json", "w", encoding="utf-8") as f:
+                json.dump({"pages": all_pages}, f, ensure_ascii=False, indent=2)
+            print(f"  -> сохранено ({i+1} страниц)")
+    
+    doc.close()
+    print(f"\n=== Готово. Результат в {out_dir} ===")
+
+
+def _run_rapidocr(img_path: Path):
     res = engine(img_path)
     ocr_lines = []
     if res and res[0] is not None:
@@ -48,26 +165,17 @@ def process_pdf(pdf_path: Path, out_dir: Path):
                 "confidence": float(score),
                 "bbox": box
             })
-        
+    return ocr_lines
-        all_pages.append({
-            "page_number": i + 1,
-            "image": str(img_path.name),
-            "pdf_text_layer": raw_text,
-            "ocr_lines": ocr_lines,
-            "ocr_line_count": len(ocr_lines)
-        })
-        print(f"OCR строк: {len(ocr_lines)}")
-        
-        if (i + 1) % BATCH_SIZE == 0 or i == total - 1:
-            with open(out_dir / "full_ocr_results.json", "w", encoding="utf-8") as f:
-                json.dump({"pages": all_pages}, f, ensure_ascii=False, indent=2)
-            print(f"  -> промежуточное сохранение ({i+1} страниц)")
-    
-    doc.close()
-    print(f"\n=== Готово. Результат в {out_dir} ===")
 
 # ------------------------------------------------------------------
 def main():
+    use_qwen = "--use-qwen" in sys.argv
+    use_tiling = "--use-tiling" in sys.argv
+    if use_qwen:
+        sys.argv.remove("--use-qwen")
+    if use_tiling:
+        sys.argv.remove("--use-tiling")
+    
     if len(sys.argv) < 2:
         pdf_file = "123.pdf"
         out_name = "output_123"
@@ -82,7 +190,7 @@ def main():
         print(f"[ERR] Файл не найден: {pdf_path}")
         sys.exit(1)
     
-    process_pdf(pdf_path, out_dir)
+    process_pdf(pdf_path, out_dir, use_qwen=use_qwen, use_tiling=use_tiling)
 
 if __name__ == "__main__":
     main()

rag_query.py

@@ -83,7 +83,8 @@ def get_lmstudio_backend(model: str = "qwen2.5:14b"):
             temperature=kwargs.get("temperature", 0.3),
             max_tokens=kwargs.get("max_tokens", 1024),
         )
-        return response.choices[0].message.content
+        content = response.choices[0].message.content
+        return content if content is not None else ""
 
     embed_model = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
 
@@ -117,7 +118,8 @@ def get_opencode_backend(model: str = "nemotron-3-super-free"):
             temperature=kwargs.get("temperature", 0.3),
             max_tokens=kwargs.get("max_tokens", 1024),
         )
-        return response.choices[0].message.content
+        content = response.choices[0].message.content
+        return content if content is not None else ""
 
     embed_model = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
 

test_qwen_ocr.py

@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Тест Alibaba Cloud DashScope qwen-vl-ocr на чертеже.
+
+Использование:
+    python test_qwen_ocr.py <png_file>
+"""
+
+import os
+import sys
+import base64
+import json
+from pathlib import Path
+from openai import OpenAI
+
+# Загрузить ключ из .env (рядом со скриптом)
+env_path = Path(__file__).parent / ".env"
+API_KEY = None
+if env_path.exists():
+    for line in env_path.read_text().splitlines():
+        if line.startswith("DASHSCOPE_API_KEY="):
+            API_KEY = line.split("=", 1)[1].strip()
+            os.environ["DASHSCOPE_API_KEY"] = API_KEY
+            break
+
+if not API_KEY:
+    API_KEY = os.environ.get("DASHSCOPE_API_KEY")
+BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
+MODEL = "qwen-vl-ocr"
+
+
+def encode_image(image_path: Path) -> str:
+    with open(image_path, "rb") as f:
+        return base64.b64encode(f.read()).decode("utf-8")
+
+
+def test_ocr(image_path: Path):
+    client = OpenAI(api_key=API_KEY, base_url=BASE_URL)
+    
+    b64 = encode_image(image_path)
+    data_url = f"data:image/png;base64,{b64}"
+    
+    print(f"Отправляем {image_path.name} в qwen-vl-ocr...")
+    print(f"Размер файла: {image_path.stat().st_size / 1024 / 1024:.1f} MB")
+    
+    response = client.chat.completions.create(
+        model=MODEL,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": (
+                            "Распознай все текстовые элементы на этом чертеже. "
+                            "Для каждого текста укажи:\n"
+                            "- сам текст\n"
+                            "- координаты bbox (x1,y1,x2,y2)\n"
+                            "- confidence (если доступен)\n"
+                            "Ответь в формате JSON-массива."
+                        ),
+                    },
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                ],
+            }
+        ],
+        temperature=0.1,
+        max_tokens=2048,
+    )
+    
+    raw = response.choices[0].message.content
+    print("\n=== ОТВЕТ МОДЕЛИ ===")
+    print(raw[:2000])
+    print("=" * 50)
+    
+    # Сохранить результат
+    out_path = image_path.parent / f"qwen_ocr_result_{image_path.stem}.json"
+    with open(out_path, "w", encoding="utf-8") as f:
+        f.write(raw)
+    print(f"\n[OK] Сохранено: {out_path}")
+
+
+def describe_image(image_path: Path):
+    """Просто описание того, что модель видит на чертеже."""
+    client = OpenAI(api_key=API_KEY, base_url=BASE_URL)
+    
+    b64 = encode_image(image_path)
+    data_url = f"data:image/png;base64,{b64}"
+    
+    print(f"\nОтправляем {image_path.name} на описание...")
+    
+    response = client.chat.completions.create(
+        model=MODEL,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {
+                        "type": "text",
+                        "text": (
+                            "Опиши подробно, что ты видишь на этом изображении. "
+                            "Чертеж здания или что-то другое? Какие элементы видны? "
+                            "Размеры, текст, линии, оси — всё, что различимо."
+                        ),
+                    },
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                ],
+            }
+        ],
+        temperature=0.3,
+        max_tokens=1024,
+    )
+    
+    desc = response.choices[0].message.content
+    print("\n=== ОПИСАНИЕ ===")
+    print(desc)
+    print("=" * 50)
+    return desc
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: python test_qwen_ocr.py <png_file> [--describe]")
+        sys.exit(1)
+    
+    image_path = Path(sys.argv[1])
+    if not image_path.exists():
+        print(f"[ERR] Файл не найден: {image_path}")
+        sys.exit(1)
+    
+    if "--describe" in sys.argv:
+        describe_image(image_path)
+    else:
+        test_ocr(image_path)
+
+
+if __name__ == "__main__":
+    main()

tiling_ocr.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Tiling OCR для больших чертежей.
+
+Разрезает PNG на перекрывающиеся кропы, прогоняет OCR на каждом,
+объединяет результаты с дедупликацией.
+
+Эффект: каждый кроп масштабирован "крупнее" для OCR — мелкий текст
+находится на бОльшем % площади кропа.
+"""
+
+import sys
+import json
+import re
+from pathlib import Path
+from typing import List, Dict, Tuple
+from PIL import Image
+from rapidocr_onnxruntime import RapidOCR
+
+
+def make_tiles(img: Image.Image, tile_size: int = 2000, overlap: int = 200) -> List[Tuple[int, int, Image.Image]]:
+    """
+    Генерирует кропы с перекрытием.
+    Возвращает: [(offset_x, offset_y, cropped_image), ...]
+    """
+    w, h = img.size
+    tiles = []
+    step = tile_size - overlap
+    
+    for y in range(0, h, step):
+        for x in range(0, w, step):
+            x2 = min(x + tile_size, w)
+            y2 = min(y + tile_size, h)
+            crop = img.crop((x, y, x2, y2))
+            tiles.append((x, y, crop))
+    
+    return tiles
+
+
+def iou_bbox(a: List, b: List) -> float:
+    """IoU двух bbox в формате [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]."""
+    def _get_rect(box):
+        if isinstance(box[0], list):
+            xs = [p[0] for p in box]
+            ys = [p[1] for p in box]
+            return min(xs), min(ys), max(xs), max(ys)
+        else:
+            return box[0], box[1], box[2], box[3]
+    
+    ax1, ay1, ax2, ay2 = _get_rect(a)
+    bx1, by1, bx2, by2 = _get_rect(b)
+    
+    ix1 = max(ax1, bx1)
+    iy1 = max(ay1, by1)
+    ix2 = min(ax2, bx2)
+    iy2 = min(ay2, by2)
+    
+    if ix2 <= ix1 or iy2 <= iy1:
+        return 0.0
+    
+    inter = (ix2 - ix1) * (iy2 - iy1)
+    area_a = (ax2 - ax1) * (ay2 - ay1)
+    area_b = (bx2 - bx1) * (by2 - by1)
+    union = area_a + area_b - inter
+    return inter / union if union > 0 else 0.0
+
+
+def run_tiling_ocr(png_path: Path, tile_size: int = 2000, overlap: int = 200, conf_threshold: float = 0.5):
+    """Основная функция."""
+    print(f"[INFO] Загрузка {png_path.name}...")
+    img = Image.open(png_path)
+    print(f"[INFO] Размер: {img.size}")
+    
+    tiles = make_tiles(img, tile_size, overlap)
+    print(f"[INFO] Кропов: {len(tiles)}")
+    
+    engine = RapidOCR()
+    all_results = []
+    
+    for i, (off_x, off_y, crop) in enumerate(tiles, 1):
+        # Временно сохранить кроп
+        tmp_path = f"/tmp/tile_{i:03d}.png"
+        crop.save(tmp_path)
+        
+        print(f"  [{i}/{len(tiles)}] tile @ ({off_x}, {off_y}) size {crop.size} ...", end=" ", flush=True)
+        res = engine(tmp_path)
+        
+        tile_lines = 0
+        if res and res[0]:
+            for item in res[0]:
+                box, txt, score = item
+                if score < conf_threshold:
+                    continue
+                # Сдвинуть bbox на offset кропа
+                shifted_box = []
+                for pt in box:
+                    shifted_box.append([pt[0] + off_x, pt[1] + off_y])
+                all_results.append({
+                    "text": txt,
+                    "confidence": float(score),
+                    "bbox": shifted_box
+                })
+                tile_lines += 1
+        print(f"{tile_lines} lines")
+    
+    # Дедупликация: если два bbox пересекаются (IoU > 0.5) — оставляем тот, что с higher confidence
+    print(f"[INFO] Дедупликация {len(all_results)} строк...")
+    unique = []
+    for r in sorted(all_results, key=lambda x: -x["confidence"]):
+        is_dup = False
+        for u in unique:
+            if iou_bbox(r["bbox"], u["bbox"]) > 0.5:
+                is_dup = True
+                break
+        if not is_dup:
+            unique.append(r)
+    
+    print(f"[OK] Уникальных строк: {len(unique)}")
+    return unique
+
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: python tiling_ocr.py <png> [tile_size] [overlap]")
+        sys.exit(1)
+    
+    png_path = Path(sys.argv[1])
+    tile_size = int(sys.argv[2]) if len(sys.argv) > 2 else 2000
+    overlap = int(sys.argv[3]) if len(sys.argv) > 3 else 200
+    
+    results = run_tiling_ocr(png_path, tile_size, overlap)
+    
+    # Сохранить результаты
+    out_json = png_path.parent / f"{png_path.stem}_tiling_ocr.json"
+    with open(out_json, "w", encoding="utf-8") as f:
+        json.dump({
+            "source": str(png_path),
+            "tile_size": tile_size,
+            "overlap": overlap,
+            "total_lines": len(results),
+            "lines": results
+        }, f, ensure_ascii=False, indent=2)
+    print(f"[OK] Сохранено: {out_json}")
+    
+    # Вывести числа
+    nums = [r for r in results if re.match(r'^\d+([,.]\d+)?$', r["text"].strip())]
+    print(f"\nНайдено {len(nums)} чисел:")
+    for n in sorted(nums, key=lambda x: x["bbox"][0][1]):
+        bbox = n["bbox"]
+        cx = sum(p[0] for p in bbox) / len(bbox)
+        cy = sum(p[1] for p in bbox) / len(bbox)
+        print(f"  {n['text']:>10}  x={cx:>8.0f}  y={cy:>8.0f}  conf={n['confidence']:.2f}")
+
+
+if __name__ == "__main__":
+    main()

visualize_dimensions.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Визуализация найденных размерных чисел на PNG.
+Рисует bbox вокруг чисел, извлечённых из OCR.
+"""
+
+import sys
+import json
+import re
+from pathlib import Path
+from PIL import Image, ImageDraw
+
+
+def visualize_dimensions(ocr_json_path: Path, png_path: Path, out_path: Path):
+    """Рисует bbox вокруг чисел на PNG."""
+    ocr = json.loads(ocr_json_path.read_text(encoding="utf-8"))
+    img = Image.open(png_path)
+    draw = ImageDraw.Draw(img)
+    
+    found = 0
+    for page in ocr.get("pages", []):
+        for line in page.get("ocr_lines", []):
+            txt = line["text"].strip()
+            if re.match(r'^\d+([,.]\d+)?$', txt):
+                bbox = line.get("bbox")
+                if bbox:
+                    # bbox: [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]
+                    if isinstance(bbox[0], list):
+                        pts = [(p[0], p[1]) for p in bbox]
+                    else:
+                        pts = [(bbox[0], bbox[1]), (bbox[2], bbox[1]), 
+                               (bbox[2], bbox[3]), (bbox[0], bbox[3])]
+                    draw.polygon(pts, outline="red", width=3)
+                    # Подпись
+                    x = min(p[0] for p in pts)
+                    y = min(p[1] for p in pts)
+                    draw.text((x, y-20), txt, fill="red")
+                    found += 1
+    
+    img.save(out_path)
+    print(f"[OK] Найдено {found} размерных чисел. Сохранено: {out_path}")
+
+
+if __name__ == "__main__":
+    if len(sys.argv) < 3:
+        print("Usage: python visualize_dimensions.py <ocr_json> <png>")
+        sys.exit(1)
+    
+    ocr_json = Path(sys.argv[1])
+    png = Path(sys.argv[2])
+    out = png.parent / f"{png.stem}_dims.png"
+    
+    visualize_dimensions(ocr_json, png, out)

vlm_describer.py

@@ -1,53 +1,139 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 """
-Генерация текстовых описаний PNG-страниц через VLM в LM Studio.
+VLM Describer — объективное извлечение структуры чертежа.
 
-Требования:
+Отправляет PNG в qwen-vl-plus (DashScope API) с промптом на фактическое
-    - Запущен LM Studio с загруженной моделью (например, qwen3-vl-4b)
+описание содержимого. НЕ ищет ошибки, НЕ оценивает качество.
-    - Сервер: http://127.0.0.1:1234/v1
+
+Результат: <output_folder>/vlm_extraction.json — структурированное описание
+каждой страницы для использования в RAG и cross-verification.
 
 Использование:
-    python vlm_describer.py <output_folder> [--prompt "..."] [--model MODEL]
+    python vlm_describer.py <output_folder> [--model MODEL]
 
-Результат: <output_folder>/vlm_descriptions.json
+Требует DASHSCOPE_API_KEY в .env или окружении.
 """
 
 import os
 import sys
 import json
 import base64
-import argparse
+import io
+import re
 from pathlib import Path
+from typing import List, Dict, Tuple
+from PIL import Image
 from openai import OpenAI
 
 # ------------------------------------------------------------------
-# Конфигурация LM Studio
+# Конфигурация
 # ------------------------------------------------------------------
-LMSTUDIO_URL = os.environ.get("LMSTUDIO_URL", "http://127.0.0.1:1234/v1")
+API_KEY = None
-LMSTUDIO_KEY = os.environ.get("LMSTUDIO_API_KEY", "lm-studio")
+BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
+DEFAULT_MODEL = "qwen-vl-plus"
 
-DEFAULT_PROMPT = (
+
-    "Опиши этот чертеж подробно. Укажи:\n"
+def _load_api_key():
-    "- Какой это этаж (если видно)\n"
+    global API_KEY
-    "- Какие оси обозначены\n"
+    if API_KEY:
-    "- Какие размеры указаны\n"
+        return API_KEY
-    "- Какие помещения/квартиры видны\n"
+    env_candidates = [
-    "- Общую компоновку и заметные детали.\n"
+        Path(__file__).parent / ".env",
-    "Отвечай на русском языке."
+        Path(__file__).parent.parent / ".env",
+    ]
+    for env_path in env_candidates:
+        if env_path.exists():
+            for line in env_path.read_text().splitlines():
+                if line.startswith("DASHSCOPE_API_KEY="):
+                    API_KEY = line.split("=", 1)[1].strip()
+                    os.environ["DASHSCOPE_API_KEY"] = API_KEY
+                    return API_KEY
+    API_KEY = os.environ.get("DASHSCOPE_API_KEY")
+    return API_KEY
+
+
+EXTRACTION_PROMPT = (
+    "Ты — система распознавания чертежей. Опиши объективно, что изображено на этой странице. "
+    "НЕ ищи ошибки, НЕ оценивай качество. Просто перечисли факты.\n\n"
+    "Ответь СТРОГО в формате JSON (без markdown):\n"
+    "{\n"
+    '  "page_type": "plan / section / elevation / specification / detail / general_view / table / unknown",\n'
+    '  "title": "заголовок или null",\n'
+    '  "beams": ["Балка Б-1"],\n'
+    '  "positions": ["П-1"],\n'
+    '  "gosts": ["ГОСТ ..."],\n'
+    '  "description": "2-3 предложения о содержимом"\n'
+    "}\n\n"
+    "ПРАВИЛА:\n"
+    "- Только реальные элементы с чертежа, не придумывай\n"
+    "- Пустой массив [] если нет элементов данного типа\n"
+    "- НЕ включай массы из таблиц в размеры\n"
+    "- Описание — только факты, без оценок"
 )
 
-client = OpenAI(base_url=LMSTUDIO_URL, api_key=LMSTUDIO_KEY)
 
+def resize_image(image_path: Path, max_size: int = 2048) -> Tuple[str, float, Tuple[int, int]]:
+    img = Image.open(image_path)
+    orig_w, orig_h = img.size
     
-def encode_image(image_path: Path) -> str:
+    if max(orig_w, orig_h) <= max_size:
         with open(image_path, "rb") as f:
-        return base64.b64encode(f.read()).decode("utf-8")
+            b64 = base64.b64encode(f.read()).decode("utf-8")
+        return b64, 1.0, (orig_w, orig_h)
+    
+    scale = max_size / max(orig_w, orig_h)
+    new_w = int(orig_w * scale)
+    new_h = int(orig_h * scale)
+    img_resized = img.resize((new_w, new_h), Image.LANCZOS)
+    
+    buf = io.BytesIO()
+    img_resized.save(buf, format="PNG")
+    b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
+    
+    return b64, scale, (orig_w, orig_h)
 
 
-def describe_image(image_path: Path, model: str, prompt: str) -> str:
+def parse_json_response(text: str) -> Dict:
-    """Отправляет PNG в VLM и получает текстовое описание."""
+    """Парсит JSON из ответа VLM."""
-    b64 = encode_image(image_path)
+    text = text.strip()
+    if text.startswith("```"):
+        text = re.sub(r"^```[a-zA-Z]*\n?", "", text)
+        text = re.sub(r"\n?```$", "", text)
+        text = text.strip()
+    
+    json_match = re.search(r'\{[\s\S]*\}', text)
+    if json_match:
+        text = json_match.group(0)
+    
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError as e:
+        print(f"[WARN] Не удалось распарсить JSON: {e}")
+        print(f"[WARN] Raw preview: {text[:500]}")
+        return {
+            "page_type": "unknown",
+            "title": None,
+            "elements": [],
+            "beams": [],
+            "positions": [],
+            "dimensions": [],
+            "gosts": [],
+            "tables": [],
+            "description": text[:500] if text else "",
+            "parse_error": str(e)
+        }
+
+
+def describe_page(image_path: Path, model: str) -> Dict:
+    """Отправляет PNG в qwen-vl API, получает структурированное описание."""
+    api_key = _load_api_key()
+    if not api_key:
+        raise RuntimeError("DASHSCOPE_API_KEY not found in .env or environment")
+    
+    client = OpenAI(api_key=api_key, base_url=BASE_URL)
+    
+    b64, scale, (orig_w, orig_h) = resize_image(image_path, max_size=2048)
     data_url = f"data:image/png;base64,{b64}"
     
     response = client.chat.completions.create(
@@ -56,58 +142,86 @@ def describe_image(image_path: Path, model: str, prompt: str) -> str:
             {
                 "role": "user",
                 "content": [
-                    {"type": "text", "text": prompt},
+                    {"type": "text", "text": EXTRACTION_PROMPT},
                     {"type": "image_url", "image_url": {"url": data_url}},
                 ],
             }
         ],
-        temperature=0.3,
+        temperature=0.1,  # низкая температура — меньше галлюцинаций
-        max_tokens=512,  # 4B модель быстро устаёт, не гоним длину
+        max_tokens=8192,
     )
-    return response.choices[0].message.content.strip()
+    raw = response.choices[0].message.content.strip()
+    
+    # Сохранить raw для отладки
+    debug_path = image_path.parent / f"{image_path.stem}_vlm_raw.txt"
+    debug_path.write_text(raw, encoding="utf-8")
+    
+    result = parse_json_response(raw)
+    
+    result["_meta"] = {
+        "image": image_path.name,
+        "original_size": [orig_w, orig_h],
+        "scale": scale,
+    }
+    
+    return result
 
 
-def process_folder(folder: Path, model: str, prompt: str):
+def run_vlm_describer(folder: Path, model: str = DEFAULT_MODEL):
-    """Обрабатывает все PNG в папке и сохраняет описания."""
+    """Запускает VLM Describer для всех PNG в папке."""
     png_files = sorted(folder.glob("page_*.png"))
     if not png_files:
         print(f"[ERR] В папке {folder} не найдены page_*.png")
         sys.exit(1)
 
-    out_path = folder / "vlm_descriptions.json"
+    out_path = folder / "vlm_extraction.json"
-    descriptions = {}
+    extractions = {}
 
-    print(f"[INFO] Найдено {len(png_files)} изображений")
+    print(f"[INFO] VLM Describer: {len(png_files)} страниц")
-    print(f"[INFO] LM Studio: {LMSTUDIO_URL}")
+    print(f"[INFO] API: DashScope ({BASE_URL})")
     print(f"[INFO] Модель: {model}\n")
 
     for i, png in enumerate(png_files, 1):
         print(f"[{i}/{len(png_files)}] {png.name} ...", end=" ", flush=True)
         try:
-            desc = describe_image(png, model, prompt)
+            data = describe_page(png, model)
-            descriptions[png.name] = desc
+            extractions[png.name] = data
-            print(f"OK ({len(desc)} chars)")
+            elem_count = len(data.get("beams", [])) + len(data.get("positions", [])) + len(data.get("gosts", []))
+            print(f"OK ({elem_count} элементов)")
         except Exception as e:
             print(f"ERR: {e}")
-            descriptions[png.name] = f"[ERROR] {e}"
+            extractions[png.name] = {
+                "page_type": "unknown",
+                "error": str(e),
+                "elements": [],
+                "beams": [],
+                "positions": [],
+                "dimensions": [],
+                "gosts": [],
+                "tables": [],
+                "description": ""
+            }
 
     with open(out_path, "w", encoding="utf-8") as f:
-        json.dump(descriptions, f, ensure_ascii=False, indent=2)
+        json.dump(extractions, f, ensure_ascii=False, indent=2)
 
-    print(f"\n[OK] Сохранено: {out_path}")
+    total_elems = sum(
+        len(v.get("beams", [])) + len(v.get("positions", [])) + len(v.get("gosts", []))
+        for v in extractions.values()
+    )
+    print(f"\n[OK] VLM extraction сохранён: {out_path}")
+    print(f"     Страниц: {len(png_files)}, Всего элементов: {total_elems}")
 
 
 def main():
-    parser = argparse.ArgumentParser(description="VLM-описания PNG через LM Studio")
+    import argparse
+    parser = argparse.ArgumentParser(description="VLM Describer для чертежей")
     parser.add_argument("folder", help="Папка с page_*.png")
-    parser.add_argument("--model", default="qwen/qwen3-vl-4b",
+    parser.add_argument("--model", default=DEFAULT_MODEL, help="Имя модели (default: qwen-vl-plus)")
-                        help="Имя модели в LM Studio (default: qwen/qwen3-vl-4b)")
-    parser.add_argument("--prompt", default=DEFAULT_PROMPT,
-                        help="Промпт для VLM")
     args = parser.parse_args()
 
     folder = Path(args.folder)
-    process_folder(folder, args.model, args.prompt)
+    run_vlm_describer(folder, args.model)
 
 
 if __name__ == "__main__":

vlm_qc_checker.py

@@ -0,0 +1,234 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+VLM-based Quality Control checker for blueprints через Alibaba Cloud API.
+
+Отправляет каждую страницу PNG в qwen-vl-plus (DashScope API)
+с промптом, просящим найти проблемы качества чертежа.
+
+Результат: <output_folder>/vlm_qc_report.json — тот же формат,
+что и dimension_qc_report.json, для совместимости с viewer.
+
+Использование:
+    python vlm_qc_checker.py <output_folder> [--model MODEL]
+
+Требует DASHSCOPE_API_KEY в .env или окружении.
+"""
+
+import os
+import sys
+import json
+import base64
+import io
+import re
+from pathlib import Path
+from typing import List, Dict, Tuple
+from PIL import Image
+from openai import OpenAI
+
+# ------------------------------------------------------------------
+# Конфигурация
+# ------------------------------------------------------------------
+API_KEY = None
+BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
+DEFAULT_MODEL = "qwen-vl-plus"  # vision model для анализа чертежей
+
+
+def _load_api_key():
+    global API_KEY
+    if API_KEY:
+        return API_KEY
+    env_candidates = [
+        Path(__file__).parent / ".env",
+        Path(__file__).parent.parent / ".env",
+    ]
+    for env_path in env_candidates:
+        if env_path.exists():
+            for line in env_path.read_text().splitlines():
+                if line.startswith("DASHSCOPE_API_KEY="):
+                    API_KEY = line.split("=", 1)[1].strip()
+                    os.environ["DASHSCOPE_API_KEY"] = API_KEY
+                    return API_KEY
+    API_KEY = os.environ.get("DASHSCOPE_API_KEY")
+    return API_KEY
+
+
+QC_PROMPT = (
+    "Ты — опытный инженер-конструктор. Проанализируй этот чертёж и найди ошибки "
+    "и проблемы в простановке размеров, расположении элементов и оформлении.\n\n"
+    "Ищи такие проблемы:\n"
+    "1. Пересечение размерных линий друг с другом\n"
+    "2. Размеры, наложенные на текст или линии\n"
+    "3. Неправильное расположение размеров (слишком близко к контуру, внутри объекта)\n"
+    "4. Пропущенные размеры (есть линии, но нет чисел)\n"
+    "5. Неправильные стрелки размеров\n"
+    "6. Размеры вне зоны видимости (слишком далеко)\n"
+    "7. Некорректные цепочки размеров (разрывы)\n"
+    "8. Плохая читаемость размеров (маленький шрифт, плохой контраст)\n\n"
+    "Ответь СТРОГО в формате JSON-массива (без markdown, без ```):\n"
+    '[\n'
+    '  {\n'
+    '    "type": "DIMENSION_OVERLAP",\n'
+    '    "severity": "warning",\n'
+    '    "message": "Описание проблемы на русском",\n'
+    '    "bbox": [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]\n'
+    '  }\n'
+    ']\n\n'
+    "Если проблем нет — верни пустой массив [].\n"
+    "severity: 'error' (критично), 'warning' (стоит исправить), 'info' (замечание).\n"
+    "bbox — координаты проблемной зоны в пикселях (если можешь определить),"
+    " иначе верни null."
+)
+
+
+def resize_image(image_path: Path, max_size: int = 2048) -> Tuple[str, float, Tuple[int, int]]:
+    """
+    Уменьшает изображение до max_size по длинной стороне для экономии токенов.
+    Возвращает (base64_string, scale_factor, (orig_w, orig_h)).
+    """
+    img = Image.open(image_path)
+    orig_w, orig_h = img.size
+    
+    if max(orig_w, orig_h) <= max_size:
+        with open(image_path, "rb") as f:
+            b64 = base64.b64encode(f.read()).decode("utf-8")
+        return b64, 1.0, (orig_w, orig_h)
+    
+    scale = max_size / max(orig_w, orig_h)
+    new_w = int(orig_w * scale)
+    new_h = int(orig_h * scale)
+    img_resized = img.resize((new_w, new_h), Image.LANCZOS)
+    
+    buf = io.BytesIO()
+    img_resized.save(buf, format="PNG")
+    b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
+    
+    return b64, scale, (orig_w, orig_h)
+
+
+def parse_vlm_response(text: str) -> list:
+    """Парсит JSON из ответа VLM."""
+    text = text.strip()
+    if text.startswith("```"):
+        text = re.sub(r"^```[a-zA-Z]*\n", "", text)
+        text = re.sub(r"\n```$", "", text)
+        text = text.strip()
+    
+    json_match = re.search(r'\[[\s\S]*\]', text)
+    if json_match:
+        text = json_match.group(0)
+    
+    try:
+        data = json.loads(text)
+        if isinstance(data, list):
+            return data
+        elif isinstance(data, dict) and "issues" in data:
+            return data["issues"]
+        else:
+            return []
+    except json.JSONDecodeError as e:
+        print(f"[WARN] Не удалось распарсить JSON: {e}")
+        print(f"[WARN] Raw text: {text[:500]}")
+        return []
+
+
+def analyze_page(image_path: Path, model: str) -> list:
+    """Отправляет PNG в qwen-vl API, получает список issues."""
+    api_key = _load_api_key()
+    if not api_key:
+        raise RuntimeError("DASHSCOPE_API_KEY not found in .env or environment")
+    
+    client = OpenAI(api_key=api_key, base_url=BASE_URL)
+    
+    b64, scale, (orig_w, orig_h) = resize_image(image_path, max_size=2048)
+    data_url = f"data:image/png;base64,{b64}"
+    
+    response = client.chat.completions.create(
+        model=model,
+        messages=[
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": QC_PROMPT},
+                    {"type": "image_url", "image_url": {"url": data_url}},
+                ],
+            }
+        ],
+        temperature=0.2,
+        max_tokens=4096,
+    )
+    raw = response.choices[0].message.content.strip()
+    
+    # Сохранить raw для отладки
+    debug_path = image_path.parent / f"{image_path.stem}_vlm_raw.txt"
+    debug_path.write_text(raw, encoding="utf-8")
+    
+    issues = parse_vlm_response(raw)
+    
+    # Масштабировать bbox обратно к оригиналу
+    if scale != 1.0:
+        for issue in issues:
+            bbox = issue.get("bbox")
+            if bbox and isinstance(bbox, list):
+                for point in bbox:
+                    if isinstance(point, list) and len(point) == 2:
+                        point[0] = round(point[0] / scale)
+                        point[1] = round(point[1] / scale)
+    
+    return issues
+
+
+def run_vlm_qc(folder: Path, model: str = DEFAULT_MODEL):
+    """Запускает VLM-QC для всех PNG в папке."""
+    png_files = sorted(folder.glob("page_*.png"))
+    if not png_files:
+        print(f"[ERR] В папке {folder} не найдены page_*.png")
+        sys.exit(1)
+
+    out_path = folder / "vlm_qc_report.json"
+    report = {"errors": [], "warnings": [], "infos": [], "source": "vlm"}
+
+    print(f"[INFO] VLM QC: {len(png_files)} страниц")
+    print(f"[INFO] API: DashScope ({BASE_URL})")
+    print(f"[INFO] Модель: {model}\n")
+
+    for i, png in enumerate(png_files, 1):
+        print(f"[{i}/{len(png_files)}] {png.name} ...", end=" ", flush=True)
+        try:
+            issues = analyze_page(png, model)
+            
+            page_num = int(png.stem.split("_")[1])
+            for issue in issues:
+                issue["page"] = page_num
+                issue["source"] = "vlm"
+                sev = issue.get("severity", "warning")
+                if sev not in ("error", "warning", "info"):
+                    sev = "warning"
+                report[f"{sev}s"].append(issue)
+            
+            print(f"OK ({len(issues)} issues)")
+        except Exception as e:
+            print(f"ERR: {e}")
+
+    with open(out_path, "w", encoding="utf-8") as f:
+        json.dump(report, f, ensure_ascii=False, indent=2)
+
+    total = sum(len(report[k]) for k in ["errors", "warnings", "infos"])
+    print(f"\n[OK] VLM QC сохранён: {out_path}")
+    print(f"     Всего замечаний: {total}")
+    print(f"     Ошибки: {len(report['errors'])}, Предупреждения: {len(report['warnings'])}, Инфо: {len(report['infos'])})")
+
+
+def main():
+    import argparse
+    parser = argparse.ArgumentParser(description="VLM QC для чертежей через qwen-vl API")
+    parser.add_argument("folder", help="Папка с page_*.png")
+    parser.add_argument("--model", default=DEFAULT_MODEL, help="Имя модели (default: qwen-vl-plus)")
+    args = parser.parse_args()
+
+    folder = Path(args.folder)
+    run_vlm_qc(folder, args.model)
+
+
+if __name__ == "__main__":
+    main()