yuezk

来围观一下

赞，我也正在用 Tauri 重写我的 VPN 项目，用的也是 MUI

总之影响因素太多了，不能奢求 API 返回的东西和官方的 Chat 一致

可能和 systemMessage 的预设和参数有关，还有就是有的壳限制了对话长度，官方是没有长度限制的。

最大的问题是中文的语料太少，可能导致模型没能真正理解学到的东西。ChatGPT 的模型是一个补全模型，它会尝试补全传给它的 prompt ，所以生成的回答往往符合语言的表达习惯（即语法没有问题），但是在准确性上不能保证，所以会经常看到它一本正经的胡说八道。

CSRF 的 token 和 sessionId 是两个东西，CSRF token 不需要长时间保存，最佳实践就是以 JS 变量或者隐藏字段的方式输出在页面中。

你如果抓一下 ChatGPT 的接口，你会发现每个 data 都会重复前面的内容，还不如 OpenAI 的接口省带宽。

```
event: add
id: 1
data: {"message":{"id":"chatcmpl-6rOOU6qBvhphYF8aL0Rw1VMitV5U8","role":"assistant","user":null,"create_time":null,"update_time":null,"end_turn":null,"weight":0,"recipient":"all","metadata":null,"content":{"content_type":"text","parts":["I"]}},"error":null}

event: add
id: 2
data: {"message":{"id":"chatcmpl-6rOOU6qBvhphYF8aL0Rw1VMitV5U8","role":"assistant","user":null,"create_time":null,"update_time":null,"end_turn":null,"weight":0,"recipient":"all","metadata":null,"content":{"content_type":"text","parts":["I'm"]}},"error":null}

event: add
id: 3
data: {"message":{"id":"chatcmpl-6rOOU6qBvhphYF8aL0Rw1VMitV5U8","role":"assistant","user":null,"create_time":null,"update_time":null,"end_turn":null,"weight":0,"recipient":"all","metadata":null,"content":{"content_type":"text","parts":["I'm sorry"]}},"error":null}

event: add
id: 4
data: {"message":{"id":"chatcmpl-6rOOU6qBvhphYF8aL0Rw1VMitV5U8","role":"assistant","user":null,"create_time":null,"update_time":null,"end_turn":null,"weight":0,"recipient":"all","metadata":null,"content":{"content_type":"text","parts":["I'm sorry,"]}},"error":null}

event: add
id: 5
data: {"message":{"id":"chatcmpl-6rOOU6qBvhphYF8aL0Rw1VMitV5U8","role":"assistant","user":null,"create_time":null,"update_time":null,"end_turn":null,"weight":0,"recipient":"all","metadata":null,"content":{"content_type":"text","parts":["I'm sorry, I"]}},"error":null}

event: add
id: 6
data: {"message":{"id":"chatcmpl-6rOOU6qBvhphYF8aL0Rw1VMitV5U8","role":"assistant","user":null,"create_time":null,"update_time":null,"end_turn":null,"weight":0,"recipient":"all","metadata":null,"content":{"content_type":"text","parts":["I'm sorry, I'm"]}},"error":null}
```

@kongkang 老哥，怎么联系？

@m16bignebula 是有关联的，这部分实际上是第三方库做的，传入 parentMessageId 之后，这个库会根据 parentMessageId 找到历史消息，然后构建一个 prompt 发给 OpenAI 的 API.

https://github.com/transitive-bullshit/chatgpt-api