1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293"""Image manipulation utilities for PDF element extraction.
Provides functions for cropping and highlighting elements on page images
using bounding box coordinates from Docling.
Also provides image compression utilities for optimizing LLM token usage
when sending images to subagents.
"""
from __future__ import annotations
import logging
from io import BytesIO
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from PIL import Image
logger = logging.getLogger(__name__)
# =============================================================================
# Image Compression Settings for LLM Subagents
# =============================================================================
# Maximum width for text verification tasks (page artifacts, footnotes, citations, lists)
# These tasks only need to verify text layout, not pixel-perfect visual formatting
TEXT_VERIFICATION_MAX_WIDTH = 800
# Maximum width for visual analysis tasks (typography)
# These tasks need higher resolution to detect bold/italic formatting
VISUAL_ANALYSIS_MAX_WIDTH = 1200
# JPEG quality for compressed images (balance between size and readability)
COMPRESSION_QUALITY = 85
# Type alias for bounding box (left, top, right, bottom)
BBox = tuple[float, float, float, float]
def crop_element(
page_image: Image.Image,
bbox: BBox,
page_width: float,
padding: int = 10,
) -> Image.Image:
"""Crop element from page image using document coordinates.
Converts Docling document coordinates to pixel coordinates and crops
the element with optional padding.
Args:
page_image: Full page image (PIL Image)
bbox: Bounding box in document coordinates (l, t, r, b)
page_width: Document page width in points (for coordinate scaling)
padding: Pixels to add around crop (default: 10)
Returns:
Cropped PIL Image containing just the element
"""
x0, y0, x1, y1 = bbox
scale = page_image.width / page_width
# Normalize coordinates (handle inverted Y from PDF coordinate system)
left = min(x0, x1)
right = max(x0, x1)
top = min(y0, y1)
bottom = max(y0, y1)
# Convert document coords to pixel coords with padding
pixel_left = max(0, int(left * scale - padding))
pixel_top = max(0, int(top * scale - padding))
pixel_right = min(page_image.width, int(right * scale + padding))
pixel_bottom = min(page_image.height, int(bottom * scale + padding))
# Final validation: ensure coordinates are valid after clamping
# This can happen when bbox is partially/fully outside the image bounds
if pixel_right <= pixel_left or pixel_bottom <= pixel_top:
import logging
logging.getLogger(__name__).warning(
f"Invalid crop box after clamping: ({pixel_left}, {pixel_top}, {pixel_right}, {pixel_bottom}), "
f"returning full page image"
)
return page_image
return page_image.crop((pixel_left, pixel_top, pixel_right, pixel_bottom))
def highlight_element(
page_image: Image.Image,
bbox: BBox,
page_width: float,
color: str = "red",
width: int = 4,
) -> Image.Image:
"""Draw highlight box around element on page image.
Creates a copy of the page image with a colored rectangle drawn
around the specified bounding box. Does not modify the original.
Args:
page_image: Full page image (PIL Image)
bbox: Bounding box in document coordinates (l, t, r, b)
page_width: Document page width in points (for coordinate scaling)
color: Rectangle color (default: "red")
width: Line width in pixels (default: 4)
Returns:
Copy of page image with highlight box drawn
"""
from PIL import ImageDraw
x0, y0, x1, y1 = bbox
img_copy = page_image.copy()
draw = ImageDraw.Draw(img_copy)
scale = img_copy.width / page_width
# Normalize coordinates (handle inverted Y from PDF coordinate system)
left = min(x0, x1)
right = max(x0, x1)
top = min(y0, y1)
bottom = max(y0, y1)
pixel_bbox = (
int(left * scale),
int(top * scale),
int(right * scale),
int(bottom * scale),
)
draw.rectangle(pixel_bbox, outline=color, width=width)
return img_copy
def crop_and_highlight(
page_image: Image.Image,
bbox: BBox,
page_width: float,
padding: int = 10,
highlight_color: str = "red",
highlight_width: int = 4,
) -> tuple[Image.Image, Image.Image]:
"""Get both cropped element and highlighted page in one call.
Convenience function for subagents that need both views:
- Cropped element for detail analysis
- Full page with highlight for context
Args:
page_image: Full page image (PIL Image)
bbox: Bounding box in document coordinates (l, t, r, b)
page_width: Document page width in points (for coordinate scaling)
padding: Pixels to add around crop (default: 10)
highlight_color: Rectangle color (default: "red")
highlight_width: Line width in pixels (default: 4)
Returns:
Tuple of (cropped_element, highlighted_page)
"""
cropped = crop_element(page_image, bbox, page_width, padding)
highlighted = highlight_element(page_image, bbox, page_width, highlight_color, highlight_width)
return cropped, highlighted
# =============================================================================
# Image Compression for LLM Token Optimization
# =============================================================================
def compress_for_text_verification(
image: "Image.Image",
max_width: int = TEXT_VERIFICATION_MAX_WIDTH,
) -> "Image.Image":
"""Compress image for text verification tasks.
Used by subagents that need to verify text layout (page artifacts,
footnotes, citations, lists) but don't need pixel-perfect visual detail.
Reduces image width to max_width while maintaining aspect ratio.
This significantly reduces token count when images are sent to LLMs.
Args:
image: Original PIL Image
max_width: Maximum width in pixels (default: 800)
Returns:
Resized PIL Image (or original if already smaller)
Example:
>>> from PIL import Image
>>> img = Image.new("RGB", (2400, 3200))
>>> compressed = compress_for_text_verification(img)
>>> compressed.size
(800, 1066) # Maintains aspect ratio
"""
if image.width <= max_width:
return image
ratio = max_width / image.width
new_height = int(image.height * ratio)
new_size = (max_width, new_height)
# Import here to avoid circular imports at module level
from PIL import Image as PILImage
resized = image.resize(new_size, PILImage.Resampling.LANCZOS)
logger.debug(
f"Compressed image for text verification: {image.size} -> {resized.size} "
f"({100 * resized.width * resized.height / (image.width * image.height):.1f}% of original pixels)"
)
return resized
def compress_for_visual_analysis(
image: "Image.Image",
max_width: int = VISUAL_ANALYSIS_MAX_WIDTH,
) -> "Image.Image":
"""Compress image for visual analysis tasks.
Used by subagents that need to detect visual formatting (typography)
where higher resolution helps identify bold/italic text.
Args:
image: Original PIL Image
max_width: Maximum width in pixels (default: 1200)
Returns:
Resized PIL Image (or original if already smaller)
"""
if image.width <= max_width:
return image
ratio = max_width / image.width
new_height = int(image.height * ratio)
new_size = (max_width, new_height)
from PIL import Image as PILImage
resized = image.resize(new_size, PILImage.Resampling.LANCZOS)
logger.debug(
f"Compressed image for visual analysis: {image.size} -> {resized.size} "
f"({100 * resized.width * resized.height / (image.width * image.height):.1f}% of original pixels)"
)
return resized
def image_to_compressed_bytes(
image: "Image.Image",
for_visual_analysis: bool = False,
format: str = "PNG",
) -> bytes:
"""Convert PIL Image to compressed bytes for LLM consumption.
Combines compression and serialization in one call. Use this when
preparing images for LLM subagents.
Args:
image: Original PIL Image
for_visual_analysis: If True, uses higher resolution (typography tasks).
If False, uses lower resolution (text verification).
format: Output format ("PNG" or "JPEG"). PNG is lossless, JPEG is smaller.
Returns:
Compressed image as bytes
Example:
>>> img = Image.open("page.png")
>>> # For footnote/citation/list/artifact detection (text verification)
>>> bytes_data = image_to_compressed_bytes(img, for_visual_analysis=False)
>>> # For typography detection (visual analysis)
>>> bytes_data = image_to_compressed_bytes(img, for_visual_analysis=True)
"""
if for_visual_analysis:
compressed = compress_for_visual_analysis(image)
else:
compressed = compress_for_text_verification(image)
buffer = BytesIO()
if format.upper() == "JPEG":
# Convert to RGB if necessary (JPEG doesn't support transparency)
if compressed.mode in ("RGBA", "LA", "P"):
compressed = compressed.convert("RGB")
compressed.save(buffer, format="JPEG", quality=COMPRESSION_QUALITY)
else:
compressed.save(buffer, format="PNG")
return buffer.getvalue()