GPT Request Body 的分析(一)
本章我们主要确认以下问题:
- 在同一个
Conversation里面. 第一次发送Message和第二次发送Message的Request Body是否相同? 后续发送的Message的的Request Body是否相同? - 调用
Image Generation时,Request Body是否相同? - 调用
Function Calling时,Request Body是否相同? - 调用
Browser Tool时,Request Body是否相同? - 在输出超长文本,单次回答无法完全输出时, 前后的
Request Body是否相同?
代码处理
我进行了总共七次问答
- 给他一段超级长的文本,让他重写(测试输出超长文本)
- 继续(继续输出超长文本没有写完的内容)
- 1-100 之和(测试 Function Calling)
- 洛杉矶天气(测试 Browser Tool)
- 绘制统计图(Function Calling)
- 对话(text completion)
- 生成图(Image Generation)
import json
import re
import logging
import os
class CompareRequest:
def __init__(self, har_file_path):
self.har_file_path = har_file_path
self.logger = logging.getLogger(__name__)
self.logger.setLevel(logging.INFO)
handler = logging.FileHandler('CompareRequest.log', mode='w', encoding='utf-8')
formatter = logging.Formatter('%(asctime)s - %(levelname)s - %(message)s')
handler.setFormatter(formatter)
self.logger.addHandler(handler)
def load_har_file(self):
with open(self.har_file_path, 'r', encoding='utf-8') as f:
self.har_data = json.load(f)
@staticmethod
def decode_unicode_escape(json_str):
return json_str.encode('utf-8').decode('unicode_escape')
def compare_headers(self):
URL_pattern = re.compile(r'^https://chatgpt.com/backend-api/conversation$')
Method_pattern = re.compile(r'^POST$')
request_bodies = []
structures = []
for entry in self.har_data['log']['entries']:
request = entry['request']
response = entry['response']
if URL_pattern.match(request['url']) and Method_pattern.match(request['method']):
if 'postData' in request:
try:
request_body = json.loads(request['postData']['text'])
request_bodies.append(request_body)
structures.append(request_body)
except json.JSONDecodeError:
self.logger.error(f"Failed to decode JSON for request URL: {request['url']}")
def get_keys(structure, parent_key=''):
keys = set()
for k, v in structure.items():
full_key = f"{parent_key}.{k}" if parent_key else k
keys.add(full_key)
if isinstance(v, dict):
keys.update(get_keys(v, full_key))
return keys
structures_keys = [get_keys(body) for body in structures]
all_keys = set().union(*structures_keys)
differences = [keys.symmetric_difference(all_keys) for keys in structures_keys]
# 创建文件夹
os.makedirs("CompareHeader", exist_ok=True)
# 将具有相同键结构的请求体分组
same_structure_groups = {}
for idx, keys in enumerate(structures_keys):
key_tuple = tuple(sorted(keys))
if key_tuple not in same_structure_groups:
same_structure_groups[key_tuple] = []
same_structure_groups[key_tuple].append(request_bodies[idx])
# 导出文件
same_structure_count = 1
different_structure_count = 1
for idx, diff in enumerate(differences):
if diff:
with open(os.path.join("CompareHeader", f"RequestBody_diff_{different_structure_count}.json"), 'w', encoding='utf-8') as f:
json.dump(request_bodies[idx], f, ensure_ascii=False, indent=2)
different_structure_count += 1
else:
with open(os.path.join("CompareHeader", f"RequestBody_same_{same_structure_count}.json"), 'w', encoding='utf-8') as f:
json.dump(request_bodies[idx], f, ensure_ascii=False, indent=2)
same_structure_count += 1
for idx, diff in enumerate(differences):
if diff:
self.logger.info(f"Request Body {idx + 1} has different structure:")
self.logger.info(diff)
else:
self.logger.info(f"Request Body {idx + 1} has the same structure as others.")
for idx, (keys, body) in enumerate(zip(structures_keys, request_bodies)):
self.logger.info(f"Request Body {idx + 1} keys:")
self.logger.info(keys)
self.logger.info("Request Body Content:")
self.logger.info(json.dumps(body, ensure_ascii=False, indent=2))
self.logger.info("-" * 80)
if __name__ == "__main__":
har_comparator = CompareRequest('SourceHar/chatgpt_firefox.har')
har_comparator.load_har_file()
har_comparator.compare_headers()处理结果
2024-07-31 22:56:19,415 - INFO - Request Body 1 has different structure:
2024-07-31 22:56:19,415 - INFO - {'conversation_id'}
2024-07-31 22:56:19,415 - INFO - Request Body 2 has the same structure as others.
2024-07-31 22:56:19,415 - INFO - Request Body 3 has the same structure as others.
2024-07-31 22:56:19,415 - INFO - Request Body 4 has the same structure as others.
2024-07-31 22:56:19,415 - INFO - Request Body 5 has the same structure as others.
2024-07-31 22:56:19,415 - INFO - Request Body 6 has the same structure as others.
2024-07-31 22:56:19,415 - INFO - Request Body 7 has the same structure as others.
2024-07-31 22:56:19,415 - INFO - Request Body 1 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'history_and_training_disabled', 'conversation_mode.kind', 'model', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa2a10e-decf-41bd-adcb-112c1b4fcb42",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"将下面的内容编写成一个博客,越长越好,越详细越好,尽可能的输出多的内容,### **Netty 常见面试题总结**\n\n很多小伙伴搞不清楚为啥要学习 Netty ，正式今天这篇文章开始之前，简单说一下自己的看法：\n\n- Netty 基于 NIO （NIO 是一种同步非阻塞的 I/O 模型，在 Java 1.4 中引入了 NIO ），使用 Netty 可以极大地简化 TCP 和 UDP 套接字服务器等网络编程，并且性能以及安全性等很多方面都非常优秀。\n- 我们平常经常接触的 Dubbo、RocketMQ、Elasticsearch、gRPC、Spark 等等热门开源项目都用到了 Netty。\n- 大部分微服务框架底层涉及到网络通信的部分都是基于 Netty 来做的，比如说 Spring Cloud 生态系统中的网关 Spring Cloud Gateway。\n\n简单总结一下和 Netty 相关问题。\n\n#### BIO,NIO 和 AIO 有啥区别？\n\n👨💻 面试官 ：先来简单介绍一下 BIO,NIO 和 AIO 3 者的区别吧！\n\n🙋 我 ：好的！"
]
},
"metadata": {},
"create_time": 1722363267.048
}
],
"parent_message_id": "aaa18741-c943-42bd-9cbe-c22948643c3e",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "40b4408c-c835-415d-a9a4-48c06a3f756f",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
2024-07-31 22:56:19,415 - INFO - Request Body 2 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'conversation_id', 'history_and_training_disabled', 'model', 'conversation_mode.kind', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa2d2c5-49a4-4695-8f91-c5c6a4b3debb",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"继续"
]
},
"metadata": {},
"create_time": 1722363328.879
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "3afab151-7ba9-491d-b8c0-350ecaf788fe",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "41931b37-b00f-4d4a-9604-a2737efdb9d8",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
2024-07-31 22:56:19,415 - INFO - Request Body 3 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'conversation_id', 'history_and_training_disabled', 'model', 'conversation_mode.kind', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa234b6-e61e-43c8-9d53-2f54be6d752c",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"使用function calling 编写1-100的和"
]
},
"metadata": {},
"create_time": 1722363391.819
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "66aa5db6-b42c-492d-89ec-849757f9d2c6",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "eb00147f-2a0f-4bde-b4c6-fc65ae8db092",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
2024-07-31 22:56:19,415 - INFO - Request Body 4 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'conversation_id', 'history_and_training_disabled', 'model', 'conversation_mode.kind', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa27199-434c-462b-b3b1-6b1eca88e6ec",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"查看现在洛杉矶的天气"
]
},
"metadata": {},
"create_time": 1722363456.823
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "8da09382-df15-4966-9b4e-e0b4d35acf89",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "00689c75-75cb-41e3-87d8-5c6233237e16",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
2024-07-31 22:56:19,415 - INFO - Request Body 5 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'conversation_id', 'history_and_training_disabled', 'model', 'conversation_mode.kind', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa2be6d-9f32-4a93-81ad-80871a903ae7",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"绘制一张统计图,这是一个抽奖系统,基础概率是0.8%,综合概率是1.8%.当80次时触发保底.绘制1-80次抽中的概率"
]
},
"metadata": {},
"create_time": 1722363487.28
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "5df564a1-dc16-42d2-bb78-64d553afdd94",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "3874976f-9894-4924-9883-cf8f6a3eafb7",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
2024-07-31 22:56:19,415 - INFO - Request Body 6 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'conversation_id', 'history_and_training_disabled', 'model', 'conversation_mode.kind', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa2de7f-3efc-4674-92d5-38054758512d",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"给我一张穿着黑色衣服的韩国美女的图片"
]
},
"metadata": {},
"create_time": 1722363555.34
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "437376be-4748-4fe7-9d4e-3b12824ef42d",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [
"如何增加抽奖中的中奖几率？",
"能否生成一个抽奖概率模拟？"
],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "37b316b0-b093-45dc-a7e2-03b0eb9461cf",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
2024-07-31 22:56:19,415 - INFO - Request Body 7 keys:
2024-07-31 22:56:19,415 - INFO - {'force_paragen', 'conversation_id', 'history_and_training_disabled', 'model', 'conversation_mode.kind', 'force_nulligen', 'reset_rate_limits', 'parent_message_id', 'force_paragen_model_slug', 'action', 'websocket_request_id', 'timezone_offset_min', 'force_rate_limit', 'force_use_sse', 'messages', 'conversation_origin', 'conversation_mode', 'suggestions'}
2024-07-31 22:56:19,415 - INFO - Request Body Content:
2024-07-31 22:56:19,415 - INFO - {
"action": "next",
"messages": [
{
"id": "aaa212b3-88d8-4069-af37-8ba91f082674",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": [
"给我一张穿着黑色皮衣的韩国美女的图片,长发,有一把伞"
]
},
"metadata": {},
"create_time": 1722363577.28
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "dbc9be6b-381b-4a4a-a7a4-bfb4e3f4adfa",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "31c1df3e-67fd-41dd-96e3-32b79a3591bf",
"force_use_sse": true,
"conversation_origin": null
}
2024-07-31 22:56:19,415 - INFO - --------------------------------------------------------------------------------
分析
通过上面的 log 文件,其实我们可以发现,在同一个Conversation里面.不论需要的工具,不论是对话,还是生成图片,或者说是function calling等各种功能.只有第一次Message的请求体不同,在发送第一次Message之后,会生成一个conversation_id,后续的Message请求,都有含有conversation_id,其他的和第一次请求体一样.
第一次的请求体结构
{
"action": "next",
"messages": [
{
"id": "aaa2a10e-decf-41bd-adcb-112c1b4fcb42",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": ["询问内容"]
},
"metadata": {},
"create_time": 1722363267.048
}
],
"parent_message_id": "aaa18741-c943-42bd-9cbe-c22948643c3e",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "40b4408c-c835-415d-a9a4-48c06a3f756f",
"force_use_sse": true,
"conversation_origin": null
}
后续的请求体结构
{
"action": "next",
"messages": [
{
"id": "aaa2d2c5-49a4-4695-8f91-c5c6a4b3debb",
"author": {
"role": "user"
},
"content": {
"content_type": "text",
"parts": ["继续"]
},
"metadata": {},
"create_time": 1722363328.879
}
],
"conversation_id": "aef1bc8f-8cbd-4b84-ad19-5960ab1e7adf",
"parent_message_id": "3afab151-7ba9-491d-b8c0-350ecaf788fe",
"model": "gpt-4o",
"timezone_offset_min": -480,
"suggestions": [],
"history_and_training_disabled": false,
"conversation_mode": {
"kind": "primary_assistant"
},
"force_paragen": false,
"force_paragen_model_slug": "",
"force_nulligen": false,
"force_rate_limit": false,
"reset_rate_limits": false,
"websocket_request_id": "41931b37-b00f-4d4a-9604-a2737efdb9d8",
"force_use_sse": true,
"conversation_origin": null
}