feat: PaddleOCR Remote-Engine (PP-OCRv5 Latin auf Hetzner x86_64)

PaddleOCR als neue engine=paddle Option in der OCR-Pipeline.
Microservice auf Hetzner (paddleocr-service/), async HTTP-Client
(paddleocr_remote.py), Frontend-Dropdown, automatisch words_first.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

admin-lehrer/components/ocr-pipeline/StepWordRecognition.tsx

@@ -60,7 +60,7 @@ export function StepWordRecognition({ sessionId, onNext, goToStep, skipHealGaps
   const [editedEntries, setEditedEntries] = useState<WordEntry[]>([])
   const [editedCells, setEditedCells] = useState<GridCell[]>([])
   const [mode, setMode] = useState<'overview' | 'labeling'>('overview')
-  const [ocrEngine, setOcrEngine] = useState<'auto' | 'tesseract' | 'rapid'>('auto')
+  const [ocrEngine, setOcrEngine] = useState<'auto' | 'tesseract' | 'rapid' | 'paddle'>('auto')
   const [usedEngine, setUsedEngine] = useState<string>('')
   const [pronunciation, setPronunciation] = useState<'british' | 'american'>('british')
   const [gridMethod, setGridMethod] = useState<'v2' | 'words_first'>('v2')
@@ -810,12 +810,13 @@ export function StepWordRecognition({ sessionId, onNext, goToStep, skipHealGaps
             {/* OCR Engine selector */}
             <select
               value={ocrEngine}
-              onChange={(e) => setOcrEngine(e.target.value as 'auto' | 'tesseract' | 'rapid')}
+              onChange={(e) => setOcrEngine(e.target.value as 'auto' | 'tesseract' | 'rapid' | 'paddle')}
               className="px-2 py-1.5 text-xs border rounded-lg dark:bg-gray-700 dark:border-gray-600"
             >
               <option value="auto">Auto (RapidOCR wenn verfuegbar)</option>
               <option value="rapid">RapidOCR (ONNX)</option>
               <option value="tesseract">Tesseract</option>
+              <option value="paddle">PaddleOCR (Hetzner)</option>
             </select>
 
             {/* Pronunciation selector (only for vocab) */}
@@ -843,6 +844,8 @@ export function StepWordRecognition({ sessionId, onNext, goToStep, skipHealGaps
               <span className={`px-2 py-0.5 rounded text-[10px] uppercase font-semibold ${
                 usedEngine === 'rapid'
                   ? 'bg-purple-100 dark:bg-purple-900/30 text-purple-700 dark:text-purple-300'
+                  : usedEngine === 'paddle'
+                  ? 'bg-blue-100 dark:bg-blue-900/30 text-blue-700 dark:text-blue-300'
                   : 'bg-gray-100 dark:bg-gray-700 text-gray-600 dark:text-gray-400'
               }`}>
                 {usedEngine}

docker-compose.yml

@@ -236,6 +236,8 @@ services:
       MINIO_BUCKET: ${MINIO_BUCKET:-breakpilot-rag}
       MINIO_SECURE: "false"
       PADDLEOCR_SERVICE_URL: http://paddleocr-service:8095
+      PADDLEOCR_REMOTE_URL: ${PADDLEOCR_REMOTE_URL:-https://hetzner.meghsakha.com:8095}
+      PADDLEOCR_API_KEY: ${PADDLEOCR_API_KEY:-}
       VAULT_ADDR: http://bp-core-vault:8200
       VAULT_TOKEN: ${VAULT_TOKEN:-breakpilot-dev-token}
       ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY:-}

docs-src/services/klausur-service/OCR-Pipeline.md

@@ -269,7 +269,7 @@ Alle Endpoints unter `/api/v1/ocr-pipeline/`.
 
 | Parameter | Default | Beschreibung |
 |-----------|---------|--------------|
-| `engine` | `auto` | OCR-Engine: `auto`, `tesseract`, `rapid` |
+| `engine` | `auto` | OCR-Engine: `auto`, `tesseract`, `rapid`, `paddle` |
 | `pronunciation` | `british` | IPA-Woerterbuch: `british` oder `american` |
 | `stream` | `false` | SSE-Streaming (nur bei `grid_method=v2`) |
 | `skip_heal_gaps` | `false` | Zeilen-Luecken nicht heilen (Overlay-Modus) |
@@ -706,10 +706,32 @@ Isolierte OCR einer einzelnen Zelle (Spalte × Zeile Schnittflaeche):
 1. **Crop:** Exakte Spalten- × Zeilengrenzen mit 3px internem Padding
 2. **Density-Check:** Ueberspringe leere Zellen (`dark_ratio < 0.005`)
 3. **Upscaling:** Kleine Crops (Hoehe < 80px) werden 3× vergroessert
-4. **OCR:** Engine-spezifisch (Tesseract, TrOCR, RapidOCR, LightON)
+4. **OCR:** Engine-spezifisch (Tesseract, TrOCR, RapidOCR, LightON, PaddleOCR)
 5. **Fallback:** Bei leerem Ergebnis → PSM 7 (Einzelzeile) statt PSM 6
 6. **Bereinigung:** `_clean_cell_text_lite()` (aggressives Noise-Filtering)
 
+### PaddleOCR Remote-Engine (`engine=paddle`)
+
+PaddleOCR (PP-OCRv5 Latin) laeuft als eigenstaendiger Microservice auf einem Hetzner x86_64 Server,
+da PaddlePaddle nicht auf ARM64 (Mac Mini) laeuft.
+
+```
+Mac Mini (klausur-service)          Hetzner (paddleocr-service)
+       │   HTTPS POST + Bild           │
+       │ ──────────────────────────▶    │  PP-OCRv5 Latin
+       │                                │  FastAPI (Port 8095)
+       │   JSON word_boxes              │  API-Key Auth
+       │ ◀──────────────────────────    │
+```
+
+**Besonderheiten:**
+
+- Erzwingt automatisch `grid_method=words_first` (full-page OCR, kein cell-crop)
+- Async HTTP-Client (`paddleocr_remote.py`) mit 30s Timeout
+- Koordinaten sind bereits absolut (kein content_bounds Offset noetig)
+- API-Key Authentifizierung ueber `X-API-Key` Header
+- Dateien: `paddleocr-service/main.py`, `services/paddleocr_remote.py`, `cv_ocr_engines.py:ocr_region_paddle()`
+
 ### Ablauf von `build_cell_grid_v2()`
 
 ```
@@ -1063,6 +1085,7 @@ ssh macmini "/usr/local/bin/docker compose -f /Users/benjaminadmin/Projekte/brea
 
 | Datum | Version | Aenderung |
 |-------|---------|----------|
+| 2026-03-12 | 4.4.0 | PaddleOCR Remote-Engine (`engine=paddle`): PP-OCRv5 Latin auf Hetzner x86_64. Neuer Microservice (`paddleocr-service/`), HTTP-Client (`paddleocr_remote.py`), Frontend-Dropdown-Option. Nutzt words_first Grid-Methode. |
 | 2026-03-12 | 4.3.0 | Words-First Grid Builder (`cv_words_first.py`): Bottom-up-Algorithmus clustert Tesseract word_boxes direkt zu Spalten/Zeilen/Zellen. Neuer `grid_method` Parameter im `/words` Endpoint. Frontend-Toggle in StepWordRecognition. |
 | 2026-03-10 | 4.2.0 | Rekonstruktion: Overlay-Modus mit Pixel-Wortpositionierung, 180°-Rotation, Sub-Session-Merging, usePixelWordPositions Hook, Box-Boundary-Schutz (box_ranges_inner) |
 | 2026-03-05 | 3.1.0 | Spalten: Seiten-Segmentierung an Sub-Headern, Word-Coverage Fallback, Segment-gefilterte Validierung |

klausur-service/backend/cv_ocr_engines.py

@@ -385,6 +385,51 @@ def ocr_region_lighton(img_bgr: np.ndarray, region: PageRegion) -> List[Dict[str
         return []
 
 
+# --- Remote PaddleOCR (Hetzner x86_64) ---
+
+
+async def ocr_region_paddle(
+    img_bgr: np.ndarray,
+    region: Optional["PageRegion"] = None,
+) -> List[Dict[str, Any]]:
+    """Run OCR via remote PaddleOCR service (Hetzner).
+
+    If *region* is given, crops before sending. Otherwise sends the full image.
+    Returns word dicts in the standard format (left/top in absolute coords).
+    """
+    from services.paddleocr_remote import ocr_remote_paddle
+
+    if region is not None:
+        crop = img_bgr[
+            region.y : region.y + region.height,
+            region.x : region.x + region.width,
+        ]
+        offset_x, offset_y = region.x, region.y
+    else:
+        crop = img_bgr
+        offset_x, offset_y = 0, 0
+
+    if crop.size == 0:
+        return []
+
+    # Encode as PNG
+    success, png_buf = cv2.imencode(".png", crop)
+    if not success:
+        logger.error("ocr_region_paddle: cv2.imencode failed")
+        return []
+
+    words, _w, _h = await ocr_remote_paddle(png_buf.tobytes())
+
+    # Shift coordinates to absolute image space
+    for w in words:
+        w["left"] += offset_x
+        w["top"] += offset_y
+        if region is not None:
+            w["region_type"] = region.type
+
+    return words
+
+
 # =============================================================================
 # Post-Processing: Deterministic Quality Fixes
 # =============================================================================

klausur-service/backend/ocr_pipeline_api.py

@@ -1865,7 +1865,7 @@ async def detect_words(
     """Build word grid from columns × rows, OCR each cell.
 
     Query params:
-        engine: 'auto' (default), 'tesseract', or 'rapid'
+        engine: 'auto' (default), 'tesseract', 'rapid', or 'paddle'
         pronunciation: 'british' (default) or 'american' — for IPA dictionary lookup
         stream: false (default) for JSON response, true for SSE streaming
         skip_heal_gaps: false (default). When true, cells keep exact row geometry
@@ -1874,6 +1874,11 @@ async def detect_words(
             'v2' uses pre-detected columns/rows (top-down).
             'words_first' clusters words bottom-up (no column/row detection needed).
     """
+    # PaddleOCR is full-page remote OCR → force words_first grid method
+    if engine == "paddle" and grid_method != "words_first":
+        logger.info("detect_words: engine=paddle requires words_first, overriding grid_method=%s", grid_method)
+        grid_method = "words_first"
+
     if session_id not in _cache:
         logger.info("detect_words: session %s not in cache, loading from DB", session_id)
         await _load_session_to_cache(session_id)
@@ -1993,6 +1998,14 @@ async def detect_words(
         t0 = time.time()
         img_h, img_w = dewarped_bgr.shape[:2]
 
+        # For paddle engine: run remote PaddleOCR full-page instead of Tesseract
+        if engine == "paddle":
+            from cv_ocr_engines import ocr_region_paddle
+
+            wf_word_dicts = await ocr_region_paddle(dewarped_bgr, region=None)
+            # PaddleOCR returns absolute coordinates, no content_bounds offset needed
+            cached["_paddle_word_dicts"] = wf_word_dicts
+        else:
             # Get word_dicts from cache or run Tesseract full-page
             wf_word_dicts = cached.get("_word_dicts")
             if wf_word_dicts is None:
@@ -2009,6 +2022,8 @@ async def detect_words(
 
         # Convert word coordinates to absolute image coordinates if needed
         # (detect_column_geometry returns words relative to content ROI)
+        # PaddleOCR already returns absolute coordinates — skip offset.
+        if engine != "paddle":
             content_bounds = cached.get("_content_bounds")
             if content_bounds:
                 lx, _rx, ty, _by = content_bounds
@@ -2035,7 +2050,7 @@ async def detect_words(
         is_vocab = bool(col_types & {'column_en', 'column_de'})
         n_rows = len(set(c['row_index'] for c in cells)) if cells else 0
         n_cols = len(columns_meta)
-        used_engine = "words_first"
+        used_engine = "paddle" if engine == "paddle" else "words_first"
 
         word_result = {
             "cells": cells,

klausur-service/backend/services/paddleocr_remote.py

@@ -0,0 +1,44 @@
+"""Remote PaddleOCR client — calls Hetzner service."""
+
+import logging
+import os
+from typing import Dict, List, Tuple
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+PADDLEOCR_REMOTE_URL = os.environ.get("PADDLEOCR_REMOTE_URL", "")
+PADDLEOCR_API_KEY = os.environ.get("PADDLEOCR_API_KEY", "")
+_TIMEOUT = 30.0
+
+
+async def ocr_remote_paddle(
+    image_bytes: bytes,
+    filename: str = "scan.png",
+) -> Tuple[List[Dict], int, int]:
+    """Send image to remote PaddleOCR, return (word_dicts, img_w, img_h)."""
+    if not PADDLEOCR_REMOTE_URL:
+        raise RuntimeError("PADDLEOCR_REMOTE_URL not configured")
+
+    headers = {}
+    if PADDLEOCR_API_KEY:
+        headers["X-API-Key"] = PADDLEOCR_API_KEY
+
+    async with httpx.AsyncClient(timeout=_TIMEOUT, verify=False) as client:
+        resp = await client.post(
+            f"{PADDLEOCR_REMOTE_URL.rstrip('/')}/ocr",
+            files={"file": (filename, image_bytes, "image/png")},
+            headers=headers,
+        )
+        resp.raise_for_status()
+        data = resp.json()
+
+    words = data.get("words", [])
+    logger.info(
+        "PaddleOCR remote returned %d words (img %dx%d)",
+        len(words),
+        data.get("image_width", 0),
+        data.get("image_height", 0),
+    )
+    return words, data["image_width"], data["image_height"]

klausur-service/backend/tests/test_paddleocr_remote.py

@@ -0,0 +1,101 @@
+"""Tests for the remote PaddleOCR client."""
+
+import pytest
+from unittest.mock import AsyncMock, patch, MagicMock
+
+
+SAMPLE_RESPONSE = {
+    "words": [
+        {"text": "Hello", "left": 10, "top": 20, "width": 80, "height": 30, "conf": 95.2},
+        {"text": "World", "left": 100, "top": 20, "width": 90, "height": 30, "conf": 91.0},
+    ],
+    "image_width": 640,
+    "image_height": 480,
+}
+
+
+@pytest.mark.asyncio
+async def test_ocr_remote_paddle_success():
+    """Successful OCR call returns word dicts and image dimensions."""
+    mock_response = MagicMock()
+    mock_response.json.return_value = SAMPLE_RESPONSE
+    mock_response.raise_for_status = MagicMock()
+
+    mock_client = AsyncMock()
+    mock_client.post.return_value = mock_response
+    mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+    mock_client.__aexit__ = AsyncMock(return_value=False)
+
+    with patch("services.paddleocr_remote.PADDLEOCR_REMOTE_URL", "https://example.com:8095"), \
+         patch("services.paddleocr_remote.PADDLEOCR_API_KEY", "test-key"), \
+         patch("httpx.AsyncClient", return_value=mock_client):
+
+        from services.paddleocr_remote import ocr_remote_paddle
+        words, w, h = await ocr_remote_paddle(b"fake-png-bytes", "test.png")
+
+    assert len(words) == 2
+    assert words[0]["text"] == "Hello"
+    assert words[1]["text"] == "World"
+    assert w == 640
+    assert h == 480
+
+    # Verify API key was sent
+    call_kwargs = mock_client.post.call_args
+    assert call_kwargs.kwargs["headers"]["X-API-Key"] == "test-key"
+
+
+@pytest.mark.asyncio
+async def test_ocr_remote_paddle_no_url():
+    """Raises RuntimeError when PADDLEOCR_REMOTE_URL is not configured."""
+    with patch("services.paddleocr_remote.PADDLEOCR_REMOTE_URL", ""):
+        from services.paddleocr_remote import ocr_remote_paddle
+        with pytest.raises(RuntimeError, match="PADDLEOCR_REMOTE_URL not configured"):
+            await ocr_remote_paddle(b"fake-png-bytes")
+
+
+@pytest.mark.asyncio
+async def test_ocr_remote_paddle_no_api_key():
+    """When no API key is set, no X-API-Key header is sent."""
+    mock_response = MagicMock()
+    mock_response.json.return_value = SAMPLE_RESPONSE
+    mock_response.raise_for_status = MagicMock()
+
+    mock_client = AsyncMock()
+    mock_client.post.return_value = mock_response
+    mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+    mock_client.__aexit__ = AsyncMock(return_value=False)
+
+    with patch("services.paddleocr_remote.PADDLEOCR_REMOTE_URL", "https://example.com:8095"), \
+         patch("services.paddleocr_remote.PADDLEOCR_API_KEY", ""), \
+         patch("httpx.AsyncClient", return_value=mock_client):
+
+        from services.paddleocr_remote import ocr_remote_paddle
+        words, w, h = await ocr_remote_paddle(b"fake-png-bytes")
+
+    assert len(words) == 2
+    call_kwargs = mock_client.post.call_args
+    assert "X-API-Key" not in call_kwargs.kwargs["headers"]
+
+
+@pytest.mark.asyncio
+async def test_ocr_remote_paddle_http_error():
+    """HTTP errors are raised to the caller."""
+    import httpx
+
+    mock_response = MagicMock()
+    mock_response.raise_for_status.side_effect = httpx.HTTPStatusError(
+        "401 Unauthorized", request=MagicMock(), response=MagicMock()
+    )
+
+    mock_client = AsyncMock()
+    mock_client.post.return_value = mock_response
+    mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+    mock_client.__aexit__ = AsyncMock(return_value=False)
+
+    with patch("services.paddleocr_remote.PADDLEOCR_REMOTE_URL", "https://example.com:8095"), \
+         patch("services.paddleocr_remote.PADDLEOCR_API_KEY", "wrong-key"), \
+         patch("httpx.AsyncClient", return_value=mock_client):
+
+        from services.paddleocr_remote import ocr_remote_paddle
+        with pytest.raises(httpx.HTTPStatusError):
+            await ocr_remote_paddle(b"fake-png-bytes")

paddleocr-service/Dockerfile

@@ -0,0 +1,16 @@
+FROM python:3.11-slim
+WORKDIR /app
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgl1-mesa-glx libglib2.0-0 curl \
+    && rm -rf /var/lib/apt/lists/*
+
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+COPY . .
+
+EXPOSE 8095
+HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=3 \
+    CMD curl -f http://127.0.0.1:8095/health || exit 1
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8095"]

paddleocr-service/main.py

@@ -0,0 +1,71 @@
+"""PaddleOCR Remote Service — PP-OCRv5 Latin auf x86_64."""
+
+import io
+import os
+
+import numpy as np
+from fastapi import FastAPI, File, Header, HTTPException, UploadFile
+from PIL import Image
+
+app = FastAPI(title="PaddleOCR Service")
+
+_engine = None
+API_KEY = os.environ.get("PADDLEOCR_API_KEY", "")
+
+
+def get_engine():
+    global _engine
+    if _engine is None:
+        from paddleocr import PaddleOCR
+
+        _engine = PaddleOCR(
+            lang="latin",
+            use_angle_cls=True,
+            show_log=False,
+        )
+    return _engine
+
+
+@app.get("/health")
+def health():
+    return {"status": "ok", "model": "PP-OCRv5-latin"}
+
+
+@app.post("/ocr")
+async def ocr(
+    file: UploadFile = File(...),
+    x_api_key: str = Header(default=""),
+):
+    if API_KEY and x_api_key != API_KEY:
+        raise HTTPException(status_code=401, detail="Invalid API key")
+
+    img_bytes = await file.read()
+    img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
+    img_np = np.array(img)
+
+    engine = get_engine()
+    result = engine.ocr(img_np)
+
+    words = []
+    for line in result[0] or []:
+        box, (text, conf) = line[0], line[1]
+        x_min = min(p[0] for p in box)
+        y_min = min(p[1] for p in box)
+        x_max = max(p[0] for p in box)
+        y_max = max(p[1] for p in box)
+        words.append(
+            {
+                "text": text,
+                "left": int(x_min),
+                "top": int(y_min),
+                "width": int(x_max - x_min),
+                "height": int(y_max - y_min),
+                "conf": round(conf * 100, 1),
+            }
+        )
+
+    return {
+        "words": words,
+        "image_width": img_np.shape[1],
+        "image_height": img_np.shape[0],
+    }

paddleocr-service/requirements.txt

@@ -0,0 +1,7 @@
+paddlepaddle>=3.0.0
+paddleocr>=2.9.0
+fastapi>=0.110.0
+uvicorn>=0.25.0
+python-multipart>=0.0.6
+Pillow>=10.0.0
+numpy>=1.24.0