fix(ocr): reduce memory usage for 16GB dev machines

- Surya models lazy-load on first OCR request instead of at startup
  (saves ~3-4GB idle RAM — Kraken stays eager at ~16MB)
- Process one page at a time in Surya engine (limits peak memory)
- RECOGNITION_BATCH_SIZE=1, DETECTOR_BATCH_SIZE=1 (slower but fits in RAM)
- Revert mem_limit back to 6GB (sufficient with these optimizations)
- Render DPI stays at 200

Idle memory: ~2GB (Kraken only). Peak during OCR: ~5-6GB (Surya loaded).

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

docker-compose.yml

@@ -78,14 +78,16 @@ services:
       dockerfile: Dockerfile
     container_name: archive-ocr
     restart: unless-stopped
-    mem_limit: 10g
-    memswap_limit: 10g
+    mem_limit: 6g
+    memswap_limit: 6g
     volumes:
       - ocr_models:/app/models
     environment:
       KRAKEN_MODEL_PATH: /app/models/german_kurrent.mlmodel
       OCR_CONFIDENCE_THRESHOLD: "0.3"
       OCR_CONFIDENCE_THRESHOLD_KURRENT: "0.5"
+      RECOGNITION_BATCH_SIZE: "1"
+      DETECTOR_BATCH_SIZE: "1"
     networks:
       - archive-net
     healthcheck:

ocr-service/engines/surya.py

@@ -6,13 +6,20 @@ logger = logging.getLogger(__name__)
 
 _recognition_predictor = None
 _detection_predictor = None
+_loaded = False
 
 
 def load_models():
-    """Eagerly load Surya models into memory. Called once at container startup."""
-    global _recognition_predictor, _detection_predictor
+    """Lazy-load Surya models on first use to save RAM at idle.
 
-    logger.info("Loading Surya models...")
+    Called automatically by extract_blocks(). Can also be called explicitly
+    to pre-warm if desired.
+    """
+    global _recognition_predictor, _detection_predictor, _loaded
+    if _loaded:
+        return
+
+    logger.info("Loading Surya models (lazy, first OCR request)...")
 
     from surya.foundation import FoundationPredictor
     from surya.recognition import RecognitionPredictor
@@ -21,6 +28,7 @@ def load_models():
     foundation_predictor = FoundationPredictor()
     _recognition_predictor = RecognitionPredictor(foundation_predictor)
     _detection_predictor = DetectionPredictor()
+    _loaded = True
 
     logger.info("Surya models loaded successfully")
 
@@ -28,22 +36,25 @@ def load_models():
 def extract_blocks(images: list, language: str = "de") -> list[dict]:
     """Run Surya OCR on a list of PIL images (one per page).
 
+    Processes one page at a time to limit peak memory usage.
     Returns a flat list of block dicts with pageNumber, x, y, width, height,
-    polygon, text. Coordinates are normalized to [0, 1] relative to page dimensions.
-    Surya 0.17+ returns polygon (4-point) natively on each text line.
+    polygon, text, words. Coordinates are normalized to [0, 1].
     """
+    load_models()
+
     all_blocks = []
 
-    predictions = _recognition_predictor(images, det_predictor=_detection_predictor)
+    for page_idx, image in enumerate(images):
+        page_w, page_h = image.size
 
-    for page_idx, page_pred in enumerate(predictions):
-        page_w, page_h = images[page_idx].size
+        # Process single page to limit peak memory
+        predictions = _recognition_predictor([image], det_predictor=_detection_predictor)
+        page_pred = predictions[0]
 
         for line in page_pred.text_lines:
-            bbox = line.bbox  # [x1, y1, x2, y2] in pixel coordinates
+            bbox = line.bbox
             x1, y1, x2, y2 = bbox
 
-            # Surya 0.17 provides polygon as list of (x, y) tuples (4 points, clockwise)
             polygon = None
             if hasattr(line, "polygon") and line.polygon and len(line.polygon) == 4:
                 polygon = [
@@ -51,7 +62,6 @@ def extract_blocks(images: list, language: str = "de") -> list[dict]:
                     for p in line.polygon
                 ]
 
-            # Extract word-level confidence for [unleserlich] marking
             words = []
             if hasattr(line, "words") and line.words:
                 for word in line.words:
@@ -73,4 +83,7 @@ def extract_blocks(images: list, language: str = "de") -> list[dict]:
                 "words": words,
             })
 
+        # Free page image after processing
+        del image
+
     return all_blocks

ocr-service/main.py

@@ -22,14 +22,13 @@ _models_ready = False
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    """Load all OCR models at startup before accepting requests."""
+    """Load lightweight models at startup. Surya loads lazily on first request."""
     global _models_ready
 
-    logger.info("Loading OCR models at startup...")
-    surya_engine.load_models()
+    logger.info("Loading Kraken model at startup (Surya loads lazily on first OCR request)...")
     kraken_engine.load_models()
     _models_ready = True
-    logger.info("All OCR models loaded — ready to accept requests")
+    logger.info("Startup complete — ready to accept requests")
 
     yield