NextChat icon indicating copy to clipboard operation
NextChat copied to clipboard

Published 20 hours ago • verified publisher icon ChatGPTNextWeb

[Bug] 使用非gpt.3.5的模型聊天导致无法自动总结内容

Open chengtx809 opened this issue 9 months ago • 15 comments

Bug Description

我使用了https://github.com/Harry-zklcdc/go-proxy-bingai项目部署了newbing的代理api， 使用的模型列表如下： -all,+gpt-4-turbo-preview,+gpt-4-vision-preview,+Precise-g4t,+Balanced-g4t,+Creative-g4t,+Precise-g4t-offline,+Balanced-g4t-offline,+Creative-g4t-offline,+Precise-g4t-18k,+Balanced-g4t-18k,+Creative-g4t-18k 但是无法总结聊天内容并自动写标题

Steps to Reproduce

直接聊天，使用gpt-4-turbo-preview模型

Expected Behavior

无法总结聊天内容并总结标题

Screenshots

No response

Deployment Method

  • [ ] Docker
  • [X] Vercel
  • [ ] Server

Desktop OS

No response

Desktop Browser

No response

Desktop Browser Version

No response

Smartphone Device

No response

Smartphone OS

No response

Smartphone Browser

No response

Smartphone Browser Version

No response

Additional Logs

No response

chengtx809 avatar Apr 29 '24 05:04 chengtx809

Bot detected the issue body's language is not English, translate it automatically.

Title: [Bug] Using non-gpt.3.5 model chat results in the inability to automatically summarize the content

Issues-translate-bot avatar Apr 29 '24 05:04 Issues-translate-bot

初步看了看好像是app/constant.ts把总结模型写死了的问题

export const SUMMARIZE_MODEL = "gpt-3.5-turbo";
export const GEMINI_SUMMARIZE_MODEL = "gemini-pro";

chengtx809 avatar Apr 29 '24 05:04 chengtx809

Bot detected the issue body's language is not English, translate it automatically.

At first glance, it seems that the summary model is written to death in app/constant.ts.

export const SUMMARIZE_MODEL = "gpt-3.5-turbo";
export const GEMINI_SUMMARIZE_MODEL = "gemini-pro";

Issues-translate-bot avatar Apr 29 '24 05:04 Issues-translate-bot

我提交了pr来修复这个bug： #4593

chengtx809 avatar Apr 29 '24 06:04 chengtx809

Bot detected the issue body's language is not English, translate it automatically.

I submitted a PR to fix this bug: #4588

Issues-translate-bot avatar Apr 29 '24 06:04 Issues-translate-bot

这个issue我记得当初考虑的是gpt4系列模型价格非常贵，回退到3.5，在加入gemini的时候讨论过，具体的issue一时没找到

Kosette avatar Apr 30 '24 09:04 Kosette

Bot detected the issue body's language is not English, translate it automatically.

I remember that I was thinking about this issue because the gpt4 series models were very expensive and I wanted to revert to 3.5. I discussed this issue when I joined gemini, but I couldn’t find the specific issue at the moment.

Issues-translate-bot avatar Apr 30 '24 09:04 Issues-translate-bot

这个issue我记得当初考虑的是gpt4系列模型价格非常贵，回退到3.5，在加入gemini的时候讨论过，具体的issue一时没找到

但是如果写死，我又不使用gpt-3.5的模型进行聊天，会导致无法总结：（ 可能个人观点问题吧，我认为与其看着一堆“新的聊天”还不如花多点token，哭笑不得了属于是

chengtx809 avatar Apr 30 '24 09:04 chengtx809

Bot detected the issue body's language is not English, translate it automatically.

Regarding this issue, I remember that I originally considered that the gpt4 series models were very expensive and would be rolled back to 3.5. I discussed this issue when I joined gemini, but I couldn’t find the specific issue at the moment.

But if I write it down and I don’t use the gpt-3.5 model for chatting, it will be impossible to summarize: ( Maybe it’s a matter of personal opinion. I think it’s better to spend more tokens than to watch a bunch of “new chats”. I think it’s funny to laugh or cry.

Issues-translate-bot avatar Apr 30 '24 09:04 Issues-translate-bot

目前，无法生成总结题目的情况会出现在使用 gpt-* 这类模型上， gpt-3.5-turbo 是 gpt-* 这类模型的总结模型，但是，通过环境变量禁用 gpt-3.5-turbo 之后，nextchat proxy会拒绝一切调用 gpt-3.5-turbo 的请求，因此导致无法总结。 gpt-3.5-turbo禁用的情况下，希望仍然使用 gpt-3.5-turbo 生成总结，还是使用 『当前模型』进行总结？

Bug Description

我使用了https://github.com/Harry-zklcdc/go-proxy-bingai项目部署了newbing的代理api， 使用的模型列表如下： -all,+gpt-4-turbo-preview,+gpt-4-vision-preview,+Precise-g4t,+Balanced-g4t,+Creative-g4t,+Precise-g4t-offline,+Balanced-g4t-offline,+Creative-g4t-offline,+Precise-g4t-18k,+Balanced-g4t-18k,+Creative-g4t-18k 但是无法总结聊天内容并自动写标题

Steps to Reproduce

直接聊天，使用gpt-4-turbo-preview模型

Expected Behavior

无法总结聊天内容并总结标题

Screenshots

No response

Deployment Method

  • [ ] Docker
  • [x] Vercel
  • [ ] Server

Desktop OS

No response

Desktop Browser

No response

Desktop Browser Version

No response

Smartphone Device

No response

Smartphone OS

No response

Smartphone Browser

No response

Smartphone Browser Version

No response

Additional Logs

No response

Dean-YZG avatar May 06 '24 10:05 Dean-YZG

目前，无法生成总结题目的情况会出现在使用 gpt-* 这类模型上， gpt-3.5-turbo 是 gpt-* 这类模型的总结模型，但是，通过环境变量禁用 gpt-3.5-turbo 之后，nextchat proxy会拒绝一切调用 gpt-3.5-turbo 的请求，因此导致无法总结。 gpt-3.5-turbo禁用的情况下，希望仍然使用 gpt-3.5-turbo 生成总结，还是使用 『当前模型』进行总结？

Bug Description

我使用了https://github.com/Harry-zklcdc/go-proxy-bingai项目部署了newbing的代理api， 使用的模型列表如下： -all,+gpt-4-turbo-preview,+gpt-4-vision-preview,+Precise-g4t,+Balanced-g4t,+Creative-g4t,+Precise-g4t-offline,+Balanced-g4t-offline,+Creative-g4t-offline,+Precise-g4t-18k,+Balanced-g4t-18k,+Creative-g4t-18k 但是无法总结聊天内容并自动写标题

Steps to Reproduce

直接聊天，使用gpt-4-turbo-preview模型

Expected Behavior

无法总结聊天内容并总结标题

Screenshots

No response

Deployment Method

  • [ ] Docker
  • [x] Vercel
  • [ ] Server

Desktop OS

No response

Desktop Browser

No response

Desktop Browser Version

No response

Smartphone Device

No response

Smartphone OS

No response

Smartphone Browser

No response

Smartphone Browser Version

No response

Additional Logs

No response

我个人希望使用当前模型进行总结

chengtx809 avatar May 07 '24 04:05 chengtx809

我也有这个问题。 版本： 2.11.2、2.12.3 都试过。 部署：Docker，未屏蔽gpt-3.5-turbo。 网页设置中，“自动生成标题”已勾选。未使用面具。 现象：无论使用什么模型，都无法生成标题。所有对话的标题都是“新的聊天”。 试过不同key、不同主机部署，均有该问题。（我用这项目很久了，似乎几个月前开始出现这种现象）

部署指令：

docker run -d -p 3003:3000  --restart always \
    -e BASE_URL=XXXX \
    -e OPENAI_API_KEY=XXXX \
    -e CODE=XXXX \
    -e ENABLE_BALANCE_QUERY=1 \
    -e DEFAULT_MODEL=gpt-4o \
    yidadaa/chatgpt-next-web

特例：如果故意使用不支持的模型（比如某个key不支持gpt-4o依然选该模型），使得返回报错，则可以触发标题总结。否则，正常能使用的模型，无法触发总结。

image

hiroi-sora avatar May 23 '24 08:05 hiroi-sora

Bot detected the issue body's language is not English, translate it automatically.

I also have this problem. Version: 2.11.2, 2.12.3 have been tried. Deployment: Docker, unblocked gpt-3.5-turbo. In the web page settings, "Automatically generate title" is checked. I have tried deployment with different keys and different hosts, but this problem still exists. Phenomenon: No matter what model is used, the title cannot be generated. All conversations are titled "New Chat." I have tried deploying with different keys and different hosts, but all have this problem. (I have been using this project for a long time, and it seems that this phenomenon started to occur a few months ago)

Deployment instructions:

docker run -d -p 3003:3000 --restart always \
    -e BASE_URL=XXXX \
    -e OPENAI_API_KEY=XXXX \
    -e CODE=XXXX \
    -e ENABLE_BALANCE_QUERY=1 \
    -e DEFAULT_MODEL=gpt-4o \
    yidadaa/chatgpt-next-web

Special case: If you deliberately use an unsupported model (for example, if a key does not support gpt-4o but still selects the model), causing an error to be returned, the title summary can be triggered.

image

Issues-translate-bot avatar May 23 '24 08:05 Issues-translate-bot

Related to !3916

gitsang avatar Jul 23 '24 02:07 gitsang

I believe that supporting configurable summary models is important, as enforcing a specific model will only lead to more issues (for instance, using gpt-4o (Azure) now still sends summary requests through the OpenAI API).

Additionally, using string prefix matching in this way is also very fragile, which could affect all future functionalities related to model naming. I think that, at the very least, the logic if (currentModel.startsWith("gpt")) should be removed.

我认为应该支持总结模型可配，强制模型只会带来更多的问题（比如使用 gpt-4o (Azure) 时仍然在用 openai 的 API 发送总结请求）

使用字符串匹配前缀的这种方式也非常的脆弱，这会导致后续所有与模型命名相关的功能都可能被这个匹配方式所影响。我认为无论如何，都至少应该删除 if (currentModel.startsWith("gpt")) 这段逻辑。

gitsang avatar Jul 23 '24 02:07 gitsang