Major Update for Bing - Supports latest bundle version and image analysis

Here it is, a much-needed update to this service which offers numerous functionalities that the old code was unable to deliver to us.

As you may know, ChatGPT Plus subscribers now have the opportunity to request image analysis directly from GPT within the chat bar. Bing has also integrated this feature into its chatbot. With this new code, you can now provide an image using a data URI, with all the following supported extensions: jpg, jpeg, png, and gif!

**What is a data URI and how can I provide an image to Bing?**

Just to clarify, a data URI is a method for encoding data directly into a URI (Uniform Resource Identifier). It is typically used for embedding small data objects like images, text, or other resources within web pages or documents. Data URIs are widely used in web applications.

To provide an image from your desktop and retrieve it as a data URI, you can use this code: [GitHub link](https://gist.github.com/jsocol/1089733).

Now, here is a code snippet you can use to provide images to Bing:

```python
import g4f

provider = g4f.Provider.Bing
user_message = [{"role": "user", "content": "Hi, describe this image."}]

response = g4f.ChatCompletion.create(
    model = g4f.models.gpt_4,
    provider = g4f.provider,  # Corrected the provider value
    messages = user_message,
    stream = True,
    image = "data:image/jpeg;base64,/9j/4AAQSkZJRgABAQEASABIAAD/4RiSRXhpZgAASUkqAAg..."  # Insert your full data URI image here
)

for message in response:
    print(message, flush=True, end='')
```

If you don't want to analyze the image, just do not specify the image parameter.

Regarding the implementation, the image is preprocessed within the Bing.py code, which can be resource-intensive for a server-side implementation. When using the Bing chatbot in your web browser, the image is preprocessed on your computer before being sent to the server. This preprocessing includes tasks like image rotation and compression. Although this implementation works, it would be more efficient to delegate image preprocessing to the client as it happens in reality. I will try to provide a JavaScript code for that at a later time.

As you saw, I did mention in the title that it is in Beta. The way the code is written, Bing can sometimes mess up its answers. Indeed, Bing does not really stream its responses as the other providers do. Bing sends its answers like this on each iteration:

"Hi,"
"Hi, this,"
"Hi, this is,"
"Hi, this is Bing."

Instead of sending each segment one at a time, it already adds them on each iteration. So, to simulate a normal streaming response, other contributors made the code wait for the next iteration to retrieve the newer segments and yield them. However, this method ignores something that Bing does.

Bing processes its responses in a markdown detector, which searches for links while the AI answers. If it finds a link, it saves it and waits until the AI finishes its answer to put all the found links at the very end of the answer. So if the AI is writing a link, but then on the next iteration, it finishes writing this link, it will then be deleted from the answer and appear later at the very end. Example:

"Here is your link reference ["
"Here is your link reference [^"
"Here is your link reference [^1"
"Here is your link reference [^1^"

And then the response would get stuck there because the markdown detector would have deleted this link reference in the next response and waited until the AI is finished to put it at the very end.

For this reason, I am working on an update to anticipate the markdown detector.
So please, if you guys notice any bugs with this new implementation, I would greatly appreciate it if you could report them on the issue tab of this repo. Thanks in advance, and I hope that all these explanations were clear to you!
This commit is contained in:
Luneye 2023-10-22 15:59:56 +02:00 committed by GitHub
parent f125f714fb
commit c400d02024
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -1,10 +1,16 @@
from __future__ import annotations from __future__ import annotations
import string
import random import random
import json import json
import os import os
import re
import io
import base64
import numpy as np
import uuid import uuid
import urllib.parse import urllib.parse
from PIL import Image
from aiohttp import ClientSession, ClientTimeout from aiohttp import ClientSession, ClientTimeout
from ..typing import AsyncResult, Messages from ..typing import AsyncResult, Messages
from .base_provider import AsyncGeneratorProvider from .base_provider import AsyncGeneratorProvider
@ -35,6 +41,7 @@ class Bing(AsyncGeneratorProvider):
proxy: str = None, proxy: str = None,
cookies: dict = None, cookies: dict = None,
tone: str = Tones.creative, tone: str = Tones.creative,
image: str = None,
**kwargs **kwargs
) -> AsyncResult: ) -> AsyncResult:
if len(messages) < 2: if len(messages) < 2:
@ -46,7 +53,7 @@ class Bing(AsyncGeneratorProvider):
if not cookies or "SRCHD" not in cookies: if not cookies or "SRCHD" not in cookies:
cookies = default_cookies cookies = default_cookies
return stream_generate(prompt, tone, context, proxy, cookies) return stream_generate(prompt, tone, image, context, proxy, cookies)
def create_context(messages: Messages): def create_context(messages: Messages):
context = "".join(f"[{message['role']}](#message)\n{message['content']}\n\n" for message in messages) context = "".join(f"[{message['role']}](#message)\n{message['content']}\n\n" for message in messages)
@ -54,14 +61,14 @@ def create_context(messages: Messages):
return context return context
class Conversation(): class Conversation():
def __init__(self, conversationId: str, clientId: str, conversationSignature: str) -> None: def __init__(self, conversationId: str, clientId: str, conversationSignature: str, imageInfo: dict=None) -> None:
self.conversationId = conversationId self.conversationId = conversationId
self.clientId = clientId self.clientId = clientId
self.conversationSignature = conversationSignature self.conversationSignature = conversationSignature
self.imageInfo = imageInfo
async def create_conversation(session: ClientSession, proxy: str = None) -> Conversation: async def create_conversation(session: ClientSession, tone: str, image: str = None, proxy: str = None) -> Conversation:
url = 'https://www.bing.com/turing/conversation/create?bundleVersion=1.1150.3' url = 'https://www.bing.com/turing/conversation/create?bundleVersion=1.1199.4'
async with await session.get(url, proxy=proxy) as response: async with await session.get(url, proxy=proxy) as response:
data = await response.json() data = await response.json()
@ -71,8 +78,65 @@ async def create_conversation(session: ClientSession, proxy: str = None) -> Conv
if not conversationId or not clientId or not conversationSignature: if not conversationId or not clientId or not conversationSignature:
raise Exception('Failed to create conversation.') raise Exception('Failed to create conversation.')
conversation = Conversation(conversationId, clientId, conversationSignature, None)
return Conversation(conversationId, clientId, conversationSignature) if isinstance(image,str):
try:
config = {
"visualSearch": {
"maxImagePixels": 360000,
"imageCompressionRate": 0.7,
"enableFaceBlurDebug": 0,
}
}
is_data_uri_an_image(image)
img_binary_data = extract_data_uri(image)
is_accepted_format(img_binary_data)
img = Image.open(io.BytesIO(img_binary_data))
width, height = img.size
max_image_pixels = config['visualSearch']['maxImagePixels']
compression_rate = config['visualSearch']['imageCompressionRate']
if max_image_pixels / (width * height) < 1:
new_width = int(width * np.sqrt(max_image_pixels / (width * height)))
new_height = int(height * np.sqrt(max_image_pixels / (width * height)))
else:
new_width = width
new_height = height
try:
orientation = get_orientation(img)
except Exception:
orientation = None
new_img = process_image(orientation, img, new_width, new_height)
new_img_binary_data = compress_image_to_base64(new_img, compression_rate)
data, boundary = build_image_upload_api_payload(new_img_binary_data, conversation, tone)
headers = session.headers.copy()
headers["content-type"] = 'multipart/form-data; boundary=' + boundary
headers["referer"] = 'https://www.bing.com/search?q=Bing+AI&showconv=1&FORM=hpcodx'
headers["origin"] = 'https://www.bing.com'
async with await session.post("https://www.bing.com/images/kblob", data=data, headers=headers, proxy=proxy) as image_upload_response:
if image_upload_response.status == 200:
image_info = await image_upload_response.json()
result = {}
if image_info.get('blobId'):
result['bcid'] = image_info.get('blobId', "")
result['blurredBcid'] = image_info.get('processedBlobId', "")
if result['blurredBcid'] != "":
result["imageUrl"] = "https://www.bing.com/images/blob?bcid=" + result['blurredBcid']
elif result['bcid'] != "":
result["imageUrl"] = "https://www.bing.com/images/blob?bcid=" + result['bcid']
if config['visualSearch']["enableFaceBlurDebug"]:
result['originalImageUrl'] = "https://www.bing.com/images/blob?bcid=" + result['blurredBcid']
else:
result['originalImageUrl'] = "https://www.bing.com/images/blob?bcid=" + result['bcid']
conversation.imageInfo = result
else:
raise Exception("Failed to parse image info.")
else:
raise Exception("Failed to upload image.")
except Exception as e:
print(f"An error happened while trying to send image: {str(e)}")
return conversation
async def list_conversations(session: ClientSession) -> list: async def list_conversations(session: ClientSession) -> list:
url = "https://www.bing.com/turing/conversation/chats" url = "https://www.bing.com/turing/conversation/chats"
@ -98,37 +162,47 @@ class Defaults:
ip_address = f"13.{random.randint(104, 107)}.{random.randint(0, 255)}.{random.randint(0, 255)}" ip_address = f"13.{random.randint(104, 107)}.{random.randint(0, 255)}.{random.randint(0, 255)}"
allowedMessageTypes = [ allowedMessageTypes = [
"ActionRequest",
"Chat", "Chat",
"Context",
"Disengaged", "Disengaged",
"Progress",
"AdsQuery", "AdsQuery",
"SemanticSerp", "SemanticSerp",
"GenerateContentQuery", "GenerateContentQuery",
"SearchQuery", "SearchQuery",
"ActionRequest", # The following message types should not be added so that it does not flood with
"Context", # useless messages (such as "Analyzing images" or "Searching the web") while it's retrieving the AI response
"Progress", # "InternalSearchQuery",
"AdsQuery", # "InternalSearchResult",
"SemanticSerp", # Not entirely certain about these two, but these parameters may be used for real-time markdown rendering.
# Keeping them could potentially complicate the retrieval of the messages because link references written while
# the AI is responding would then be moved to the very end of its message.
# "RenderCardRequest",
# "RenderContentRequest"
] ]
sliceIds = [ sliceIds = [
"winmuid3tf", "wrapuxslimt5",
"osbsdusgreccf", "wrapalgo",
"ttstmout", "wraptopalgo",
"crchatrev", "st14",
"winlongmsgtf", "arankr1_1_9_9",
"ctrlworkpay", "0731ziv2s0",
"norespwtf", "voiceall",
"tempcacheread", "1015onstblg",
"temptacache", "vsspec",
"505scss0", "cacdiscf",
"508jbcars0", "909ajcopus0",
"515enbotdets0", "scpbfmob",
"5082tsports", "rwt1",
"515vaoprvs", "cacmuidarb",
"424dagslnv1s0", "sappdlpt",
"kcimgattcf", "917fluxv14",
"427startpms0", "delaygc",
"remsaconn3p",
"splitcss3p",
"sydconfigoptt"
] ]
location = { location = {
@ -173,27 +247,128 @@ class Defaults:
} }
optionsSets = [ optionsSets = [
'saharasugg',
'enablenewsfc',
'clgalileo',
'gencontentv3',
"nlu_direct_response_filter", "nlu_direct_response_filter",
"deepleo", "deepleo",
"disable_emoji_spoken_text", "disable_emoji_spoken_text",
"responsible_ai_policy_235", "responsible_ai_policy_235",
"enablemm", "enablemm",
"h3precise"
"dtappid",
"cricinfo",
"cricinfov2",
"dv3sugg", "dv3sugg",
"nojbfedge" "iyxapbing",
"iycapbing",
"h3imaginative",
"clgalileo",
"gencontentv3",
"fluxv14",
"eredirecturl"
] ]
def format_message(msg: dict) -> str: def format_message(msg: dict) -> str:
return json.dumps(msg, ensure_ascii=False) + Defaults.delimiter return json.dumps(msg, ensure_ascii=False) + Defaults.delimiter
def build_image_upload_api_payload(image_bin: str, conversation: Conversation, tone: str):
payload = {
'invokedSkills': ["ImageById"],
'subscriptionId': "Bing.Chat.Multimodal",
'invokedSkillsRequestData': {
'enableFaceBlur': True
},
'convoData': {
'convoid': "",
'convotone': tone
}
}
knowledge_request = {
'imageInfo': {},
'knowledgeRequest': payload
}
boundary="----WebKitFormBoundary" + ''.join(random.choices(string.ascii_letters + string.digits, k=16))
data = '--' + boundary + '\r\nContent-Disposition: form-data; name="knowledgeRequest"\r\n\r\n' + json.dumps(knowledge_request,ensure_ascii=False) + "\r\n--" + boundary + '\r\nContent-Disposition: form-data; name="imageBase64"\r\n\r\n' + image_bin + "\r\n--" + boundary + "--\r\n"
return data, boundary
def is_data_uri_an_image(data_uri):
try:
# Check if the data URI starts with 'data:image' and contains an image format (e.g., jpeg, png, gif)
if not re.match(r'data:image/(\w+);base64,', data_uri):
raise ValueError("Invalid data URI image.")
# Extract the image format from the data URI
image_format = re.match(r'data:image/(\w+);base64,', data_uri).group(1)
# Check if the image format is one of the allowed formats (jpg, jpeg, png, gif)
if image_format.lower() not in ['jpeg', 'jpg', 'png', 'gif']:
raise ValueError("Invalid image format (from mime file type).")
except Exception as e:
raise e
def is_accepted_format(binary_data):
try:
check = False
if binary_data.startswith(b'\xFF\xD8\xFF'):
check = True # It's a JPEG image
elif binary_data.startswith(b'\x89PNG\r\n\x1a\n'):
check = True # It's a PNG image
elif binary_data.startswith(b'GIF87a') or binary_data.startswith(b'GIF89a'):
check = True # It's a GIF image
elif binary_data.startswith(b'\x89JFIF') or binary_data.startswith(b'JFIF\x00'):
check = True # It's a JPEG image
elif binary_data.startswith(b'\xFF\xD8'):
check = True # It's a JPEG image
elif binary_data.startswith(b'RIFF') and binary_data[8:12] == b'WEBP':
check = True # It's a WebP image
# else we raise ValueError
if not check:
raise ValueError("Invalid image format (from magic code).")
except Exception as e:
raise e
def extract_data_uri(data_uri):
try:
data = data_uri.split(",")[1]
data = base64.b64decode(data)
return data
except Exception as e:
raise e
def get_orientation(data: bytes):
try:
if data[0:2] != b'\xFF\xD8':
raise Exception('NotJpeg')
with Image.open(data) as img:
exif_data = img._getexif()
if exif_data is not None:
orientation = exif_data.get(274) # 274 corresponds to the orientation tag in EXIF
if orientation is not None:
return orientation
except Exception:
pass
def process_image(orientation, img, new_width, new_height):
try:
# Initialize the canvas
new_img = Image.new("RGB", (new_width, new_height), color="#FFFFFF")
if orientation:
if orientation > 4:
img = img.transpose(Image.FLIP_LEFT_RIGHT)
if orientation == 3 or orientation == 4:
img = img.transpose(Image.ROTATE_180)
if orientation == 5 or orientation == 6:
img = img.transpose(Image.ROTATE_270)
if orientation == 7 or orientation == 8:
img = img.transpose(Image.ROTATE_90)
new_img.paste(img, (0, 0))
return new_img
except Exception as e:
raise e
def compress_image_to_base64(img, compression_rate):
try:
output_buffer = io.BytesIO()
img.save(output_buffer, format="JPEG", quality=int(compression_rate * 100))
base64_image = base64.b64encode(output_buffer.getvalue()).decode('utf-8')
return base64_image
except Exception as e:
raise e
def create_message(conversation: Conversation, prompt: str, tone: str, context: str=None) -> str: def create_message(conversation: Conversation, prompt: str, tone: str, context: str=None) -> str:
request_id = str(uuid.uuid4()) request_id = str(uuid.uuid4())
struct = { struct = {
'arguments': [ 'arguments': [
@ -213,6 +388,7 @@ def create_message(conversation: Conversation, prompt: str, tone: str, context:
'requestId': request_id, 'requestId': request_id,
'messageId': request_id, 'messageId': request_id,
}, },
"scenario": "SERP",
'tone': tone, 'tone': tone,
'spokenTextMode': 'None', 'spokenTextMode': 'None',
'conversationId': conversation.conversationId, 'conversationId': conversation.conversationId,
@ -225,7 +401,11 @@ def create_message(conversation: Conversation, prompt: str, tone: str, context:
'target': 'chat', 'target': 'chat',
'type': 4 'type': 4
} }
if conversation.imageInfo != None and "imageUrl" in conversation.imageInfo and "originalImageUrl" in conversation.imageInfo:
struct['arguments'][0]['message']['originalImageUrl'] = conversation.imageInfo['originalImageUrl']
struct['arguments'][0]['message']['imageUrl'] = conversation.imageInfo['imageUrl']
struct['arguments'][0]['experienceType'] = None
struct['arguments'][0]['attachedFileInfo'] = {"fileName": None, "fileType": None}
if context: if context:
struct['arguments'][0]['previousMessages'] = [{ struct['arguments'][0]['previousMessages'] = [{
"author": "user", "author": "user",
@ -239,6 +419,7 @@ def create_message(conversation: Conversation, prompt: str, tone: str, context:
async def stream_generate( async def stream_generate(
prompt: str, prompt: str,
tone: str, tone: str,
image: str = None,
context: str = None, context: str = None,
proxy: str = None, proxy: str = None,
cookies: dict = None cookies: dict = None
@ -248,7 +429,7 @@ async def stream_generate(
cookies=cookies, cookies=cookies,
headers=Defaults.headers, headers=Defaults.headers,
) as session: ) as session:
conversation = await create_conversation(session, proxy) conversation = await create_conversation(session, tone, image, proxy)
try: try:
async with session.ws_connect( async with session.ws_connect(
f'wss://sydney.bing.com/sydney/ChatHub', f'wss://sydney.bing.com/sydney/ChatHub',
@ -264,7 +445,6 @@ async def stream_generate(
response_txt = '' response_txt = ''
returned_text = '' returned_text = ''
final = False final = False
while not final: while not final:
msg = await wss.receive(timeout=900) msg = await wss.receive(timeout=900)
objects = msg.data.split(Defaults.delimiter) objects = msg.data.split(Defaults.delimiter)
@ -299,4 +479,4 @@ async def stream_generate(
raise Exception(f"{result['value']}: {result['message']}") raise Exception(f"{result['value']}: {result['message']}")
return return
finally: finally:
await delete_conversation(session, conversation, proxy) await delete_conversation(session, conversation, proxy)