Update app.py

app.py

@@ -8,6 +8,12 @@ Scope (intentionally small):
   - Footer band: same pattern, with light dedup so we do not paste a full
     transaction dump twice when the body already captured it
 
+Universal image pipeline (same for every PDF, no keywords / no bank logic):
+  - Higher rasterization scale + extra white padding so fine print, boxed
+    section labels, and right-aligned amounts sit farther from the clip edge.
+  - Mild contrast + unsharp mask on every raster sent to the model so
+    thin rules and small glyphs are easier to read before recognition.
+
 Explicitly omitted vs the heavy Space build:
   - No HTML table rewriting, text-layer row injection, institution-specific
     splits (UCB / Navy / TD / First Horizon / …), or doc-wide dedupe passes
@@ -55,19 +61,33 @@ log = logging.getLogger("glmocr_simple_app")
 logging.basicConfig(level=logging.INFO)
 
 # ---------------------------------------------------------------------------
-# Settings
+# Settings — tuned for dense financial PDFs; applies to every document
 # ---------------------------------------------------------------------------
 
-GLMOCR_API_KEY = "cee1d52dd91a4ab591b3f6e105f8ad89.LgbQTECuzX0zrito"
+GLMOCR_API_KEY = os.environ.get("GLMOCR_API_KEY", "").strip()
 if not GLMOCR_API_KEY:
     log.warning("GLMOCR_API_KEY is not set; GlmOcr() will fail until you export it.")
 
-RENDER_SCALE = 2.2
-PAD_LEFT_FRAC = 0.02
-PAD_RIGHT_FRAC = 0.06
-PAD_TOP_FRAC = 0.01
-PAD_BOTTOM_FRAC = 0.01
+# Rasterization: higher scale = more pixels per PDF point (helps small type,
+# boxed headers, and narrow columns). Same constant for all uploads.
+RENDER_SCALE = 2.85
+
+# White margin as a fraction of page width/height after render. Extra right
+# margin helps right-aligned currency columns that hug the page edge.
+PAD_LEFT_FRAC = 0.03
+PAD_RIGHT_FRAC = 0.10
+PAD_TOP_FRAC = 0.015
+PAD_BOTTOM_FRAC = 0.015
+
 ENABLE_CONTRAST = True
+# Slight contrast lift only; same factor for every file.
+CONTRAST_FACTOR = 1.12
+
+# Subtle edge enhancement after contrast (helps hairlines and small digits).
+ENABLE_UNSHARP = True
+UNSHARP_RADIUS = 0.85
+UNSHARP_PERCENT = 65
+UNSHARP_THRESHOLD = 2
 
 DEFAULT_ZONE_FRAC = 0.12
 PDF_HEADER_BAND_FRAC = 0.10
@@ -78,9 +98,35 @@ PDF_FOOTER_BAND_FRAC = 0.88
 MIN_CROP_HEIGHT = 112
 MIN_CROP_PIXELS = 112 * 112
 
+# PNG compression 0–9; lower = less loss before GLM-OCR (same for all PDFs).
+PAGE_PNG_COMPRESS_LEVEL = 3
+# JPEG quality for small header/footer crops sent to the API.
+ZONE_JPEG_QUALITY = 95
+
 _parser = None
 
 
+def _enhance_raster_for_ocr(img):
+    """
+    Improve legibility of every raster passed to GLM-OCR (full pages and
+    header/footer crops). No document text or keywords — same pipeline for
+    all PDFs and images.
+    """
+    from PIL import ImageEnhance, ImageFilter
+
+    if ENABLE_CONTRAST:
+        img = ImageEnhance.Contrast(img).enhance(CONTRAST_FACTOR)
+    if ENABLE_UNSHARP:
+        img = img.filter(
+            ImageFilter.UnsharpMask(
+                radius=UNSHARP_RADIUS,
+                percent=UNSHARP_PERCENT,
+                threshold=UNSHARP_THRESHOLD,
+            )
+        )
+    return img
+
+
 def get_parser():
     global _parser
     if glmocr is None:
@@ -186,7 +232,7 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
         fd, path = tempfile.mkstemp(suffix=".jpg")
         os.close(fd)
         try:
-            crop.save(path, "JPEG", quality=92)
+            crop.save(path, "JPEG", quality=ZONE_JPEG_QUALITY)
             parser = get_parser()
             out = parser.parse(path)
             if not isinstance(out, list):
@@ -308,7 +354,7 @@ def light_stabilize_markdown(page_md: str) -> str:
 
 def render_pdf_pages_to_images(pdf_path: str) -> Tuple[List[str], List[int]]:
     import pymupdf as fitz
-    from PIL import Image, ImageEnhance
+    from PIL import Image
 
     doc = fitz.open(pdf_path)
     page_images: List[str] = []
@@ -319,9 +365,7 @@ def render_pdf_pages_to_images(pdf_path: str) -> Tuple[List[str], List[int]]:
         pix = page.get_pixmap(matrix=fitz.Matrix(RENDER_SCALE, RENDER_SCALE), alpha=False)
 
         img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
-
-        if ENABLE_CONTRAST:
-            img = ImageEnhance.Contrast(img).enhance(1.12)
+        img = _enhance_raster_for_ocr(img)
 
         w, h = img.size
         pad_l = int(w * PAD_LEFT_FRAC)
@@ -335,7 +379,7 @@ def render_pdf_pages_to_images(pdf_path: str) -> Tuple[List[str], List[int]]:
             img = canvas
 
         img_path = os.path.join(tempfile.gettempdir(), f"glmocr_page_{os.getpid()}_{i}.png")
-        img.save(img_path, "PNG", compress_level=6)
+        img.save(img_path, "PNG", compress_level=PAGE_PNG_COMPRESS_LEVEL)
         page_images.append(img_path)
         page_heights.append(img.height)