perf: switch ocr to trocr large

app.py

@@ -3,12 +3,21 @@ import re
 from typing import Any, Dict, List, Optional, Sequence
 
 import gradio as gr
+import torch
 from PIL import Image, ImageDraw
 from transformers import pipeline
 
 # --- OCR pipeline ---------------------------------------------------------
-# We use a light-weight printed-text OCR model that works well for receipts/labels.
-ocr = pipeline("image-to-text", model="microsoft/trocr-base-printed")
+# Use a high-capacity OCR model for better accuracy on prescription labels.
+MODEL_ID = "microsoft/trocr-large-printed"
+
+
+def _load_ocr():
+    device = 0 if torch.cuda.is_available() else -1
+    return pipeline("image-to-text", model=MODEL_ID, device=device)
+
+
+ocr = _load_ocr()
 
 # Korean keywords describing time slots on prescription labels.
 TIME_KEYWORDS = [

requirements.txt

@@ -1,4 +1,5 @@
 transformers
+torch
 gradio
 Pillow
-torch
+torch