-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathllm.py
323 lines (273 loc) · 12.5 KB
/
llm.py
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
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
from abc import ABC, abstractmethod
from typing import Type, TypeVar
import base64
import os
import json
from doc2json import process_docx
import fitz
from PIL import Image
import io
import boto3
from botocore.config import Config
import re
from PIL import Image
import io
import math
import gradio
# constants
log_to_console = False
use_document_message_type = False # AWS document message type usage
LLMClass = TypeVar('LLMClass', bound='LLM')
class LLM:
@staticmethod
def create_llm(model: str) -> Type[LLMClass]:
return LLM()
def generate_body(self, message, history):
messages = []
# AWS API requires strict user, assi, user, ... sequence
lastTypeHuman = False
for msg in history:
if msg['role'] == "user":
if lastTypeHuman:
last_msg = messages.pop()
user_msg_parts = last_msg["content"]
else:
user_msg_parts = []
content = msg['content']
if isinstance(content, gradio.File) or isinstance(content, gradio.Image):
user_msg_parts.extend(self._process_file(content.value['path']))
elif isinstance(content, tuple):
user_msg_parts.extend(self._process_file(content[0]))
else:
user_msg_parts.extend([{"text": content}])
messages.append({"role": "user", "content": user_msg_parts})
lastTypeHuman = True
else:
messages.append({
"role": "assistant",
"content":[{"text": msg['content']}]
})
lastTypeHuman = False
if lastTypeHuman:
last_msg = messages.pop()
user_msg_parts = last_msg["content"]
else:
user_msg_parts = []
if message:
if message["text"]:
user_msg_parts.append({"text": message["text"]})
if message["files"]:
for file in message["files"]:
user_msg_parts.extend(self._process_file(file))
if user_msg_parts:
messages.append({"role": "user", "content": user_msg_parts})
return messages
def _process_file(self, file_path):
if use_document_message_type and self._is_supported_document_type(file_path):
return [self._create_document_message(file_path)]
else:
return self._encode_file(file_path)
def _is_supported_document_type(self, file_path):
supported_extensions = ['.pdf', '.csv', '.doc', '.docx', '.xls', '.xlsx', '.html', '.txt', '.md']
return os.path.splitext(file_path)[1].lower() in supported_extensions
def _create_document_message(self, file_path):
with open(file_path, 'rb') as file:
file_content = file.read()
file_name = re.sub(r'[^a-zA-Z0-9\s\-\(\)\[\]]', '', os.path.basename(file_path))[:200].strip() or "unnamed_file"
file_extension = os.path.splitext(file_path)[1][1:] # Remove the dot
return {
"document": {
"name": file_name,
"format": file_extension,
"source": {
"bytes": file_content
}
}
}
def _encode_file(self, fn: str) -> list:
if fn.endswith(".docx"):
return [{"text": process_docx(fn)}]
elif fn.endswith(".pdf"):
return self._process_pdf_img(fn)
else:
with open(fn, mode="rb") as f:
content = f.read()
if isinstance(content, bytes):
try:
# try to add as image
image_data = self._encode_image(content)
return [{"image": image_data}]
except:
# not an image, try text
content = content.decode('utf-8', 'replace')
else:
content = str(content)
fname = os.path.basename(fn)
return [{"text": f"``` {fname}\n{content}\n```"}]
def _process_pdf_img(self, pdf_fn: str):
pdf = fitz.open(pdf_fn)
message_parts = []
page_scales = {} # Cache for similar page sizes
def calculate_tokens(width, height):
return (width * height) / 750
for page in pdf.pages():
page_rect = page.rect
orig_width = page_rect.width
orig_height = page_rect.height
page_key = (orig_width, orig_height)
# Use cached scale as starting point if available
scale = page_scales.get(page_key, 1.0)
while True:
# Render with current scale
mat = fitz.Matrix(scale, scale)
pix = page.get_pixmap(matrix=mat, alpha=False)
# Check actual rendered dimensions
actual_tokens = calculate_tokens(pix.width, pix.height)
actual_long_edge = max(pix.width, pix.height)
if actual_long_edge <= 1568 and actual_tokens <= 1600:
# We found a good scale, cache it
if page_key not in page_scales:
page_scales[page_key] = scale
break
# Calculate new scale factor based on both constraints
if actual_long_edge > 1568:
scale_factor = min(1568 / actual_long_edge, 0.9)
else:
scale_factor = min(math.sqrt(1600 / actual_tokens), 0.9)
scale *= scale_factor
# Convert to PIL Image
img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
# Handle compression
quality = 95
while True:
buffer = io.BytesIO()
img.save(buffer, format="webp", quality=quality)
img_bytes = buffer.getvalue()
if len(img_bytes) <= 5 * 1024 * 1024 or quality <= 20:
break
quality = max(int(quality * 0.9), 20)
message_parts.append({"text": f"Page {page.number + 1} of file '{pdf_fn}'"})
message_parts.append({"image": {
"format": "webp",
"source": {"bytes": img_bytes}
}})
pdf.close()
return message_parts
def _encode_image(self, image_data):
try:
# Open the image using Pillow
img = Image.open(io.BytesIO(image_data))
original_format = img.format.lower()
except IOError:
raise Exception("Unknown image type")
# Ensure correct orientation based on EXIF
try:
exif = img._getexif()
if exif:
orientation = exif.get(274) # 274 is the orientation tag
if orientation:
# Rotate or flip based on EXIF orientation
if orientation == 3:
img = img.rotate(180, expand=True)
elif orientation == 6:
img = img.rotate(270, expand=True)
elif orientation == 8:
img = img.rotate(90, expand=True)
except:
pass # If EXIF processing fails, use image as-is
# check if within the limits for Claude as per https://docs.anthropic.com/en/docs/build-with-claude/vision
def calculate_tokens(width, height):
return (width * height) / 750
tokens = calculate_tokens(img.width, img.height)
long_edge = max(img.width, img.height)
format_ok = original_format in ["jpg", "jpeg", "png", "webp"]
# Check if the image already meets all requirements
if format_ok and (long_edge <= 1568 and tokens <= 1600 and len(image_data) <= 5 * 1024 * 1024):
return {
"format": original_format,
"source": {"bytes": image_data}
}
# If we need to modify the image, proceed with resizing and/or compression
orig_scale_factor = 1
orig_img = img
while long_edge > 1568 or tokens > 1600:
if long_edge > 1568:
scale_factor = min(1568 / long_edge, 0.9)
else:
scale_factor = min(math.sqrt(1600 / tokens), 0.9)
scale_factor = orig_scale_factor * scale_factor
orig_scale_factor = scale_factor
new_width = int(orig_img.width * scale_factor)
new_height = int(orig_img.height * scale_factor)
img = orig_img.resize((new_width, new_height), Image.LANCZOS)
long_edge = max(img.width, img.height)
tokens = calculate_tokens(img.width, img.height)
# Try to save in original format first
buffer = io.BytesIO()
out_fmt = "png" if original_format == "png" else "webp"
img.save(buffer, format=out_fmt, quality=95 if out_fmt == "webp" else None)
image_data = buffer.getvalue()
# If the image is still too large, switch to WebP and compress
if len(image_data) > 5 * 1024 * 1024:
quality = 95
while len(image_data) > 5 * 1024 * 1024:
quality = max(int(quality * 0.9), 20)
buffer = io.BytesIO()
img.save(buffer, format="webp", quality=quality)
image_data = buffer.getvalue()
if quality == 20:
# If we've reached quality 20 and it's still too large, resize
scale_factor = 0.9
new_width = int(img.width * scale_factor)
new_height = int(img.height * scale_factor)
img = img.resize((new_width, new_height), Image.LANCZOS)
quality = 95 # Reset quality for the resized image
return {
"format": "webp",
"source": {"bytes": image_data}
}
def read_response(self, response_stream):
"""
Handles response stream that may contain both regular text and tool use requests.
Yields tuples of (text, tool_request, stop_reason) where:
- text: accumulated text response
- tool_request: dict with tool use details if present, None otherwise
- stop_reason: string indicating why stream stopped, None while streaming
"""
message = {}
content = []
message['content'] = content
tool_use = {}
text = ''
stop_reason = None
for chunk in response_stream:
if 'messageStart' in chunk:
message['role'] = chunk['messageStart']['role']
elif 'contentBlockStart' in chunk:
tool = chunk['contentBlockStart']['start']['toolUse']
tool_use['toolUseId'] = tool['toolUseId']
tool_use['name'] = tool['name']
elif 'contentBlockDelta' in chunk:
delta = chunk['contentBlockDelta']['delta']
if 'toolUse' in delta:
if 'input' not in tool_use:
tool_use['input'] = ''
tool_use['input'] += delta['toolUse']['input']
elif 'text' in delta:
text += delta['text']
yield None, delta['text']
elif 'contentBlockStop' in chunk:
if 'input' in tool_use:
tool_use['input'] = json.loads(tool_use['input'])
content.append({'toolUse': tool_use})
tool_use = {}
else:
content.append({'text': text})
elif 'messageStop' in chunk:
stop_reason = chunk['messageStop']['stopReason']
yield stop_reason, message
elif 'metadata' in chunk and 'usage' in chunk['metadata'] and log_to_console:
print("\nToken usage:")
print(f"Input tokens: {metadata['usage']['inputTokens']}")
print(f"Output tokens: {metadata['usage']['outputTokens']}")
print(f"Total tokens: {metadata['usage']['totalTokens']}")