Use RapidOCR/EasyOCR defaults; avoid Paddle runtime downloads

requirements.txt

@@ -12,7 +12,8 @@ pydantic==2.12.5
 opencv-python-headless==4.11.0.86
 numpy==1.26.4
 ultralytics==8.4.12
-paddleocr==2.7.3
-paddlepaddle==2.6.2
+rapidocr-onnxruntime==1.4.4
+easyocr==1.7.2
 torch==2.5.1
+torchvision==0.20.1
 clip @ git+https://github.com/openai/CLIP.git

run_manager.py

@@ -479,20 +479,28 @@ def start_run(
 
     started = time.time()
     logs_path = _logs_path(run_id)
+    child_env = os.environ.copy()
+    child_env["PYTHONUNBUFFERED"] = "1"
+    child_env.setdefault("PYTHONIOENCODING", "utf-8")
+    # Prefer faster OCR initialization by default in hosted runs.
+    # Space-level env vars can override these.
+    child_env.setdefault("OCR_MODE", "cpu")
+    child_env.setdefault("OCR_BACKEND_CPU", "rapidocr")
+    child_env.setdefault("OCR_BACKEND_GPU", "easyocr")
+
     log_fh = open(logs_path, "a", encoding="utf-8", buffering=1)
     log_fh.write(
         f"[runner] run_id={run_id} variant={variant} started_at_epoch={started}\n"
         f"[runner] command={' '.join(cmd)}\n"
         f"[runner] cwd={PIPELINES_DIR}\n\n"
         f"[runner] heartbeat_interval_sec={log_heartbeat_sec}\n"
-        f"[runner] python_unbuffered=1\n\n"
+        f"[runner] python_unbuffered=1\n"
+        f"[runner] ocr_mode={child_env.get('OCR_MODE')} "
+        f"ocr_backend_cpu={child_env.get('OCR_BACKEND_CPU')} "
+        f"ocr_backend_gpu={child_env.get('OCR_BACKEND_GPU')}\n\n"
     )
     log_fh.flush()
 
-    child_env = os.environ.copy()
-    child_env["PYTHONUNBUFFERED"] = "1"
-    child_env.setdefault("PYTHONIOENCODING", "utf-8")
-
     proc = subprocess.Popen(
         cmd,
         cwd=str(PIPELINES_DIR),
@@ -578,4 +586,3 @@ def get_final_output(run_id: str, condensed: bool = False) -> Dict[str, Any]:
     if not out_file.exists():
         raise FileNotFoundError(f"Output not found: {out_file}")
     return _read_json(out_file)
-