[Bug] Model Reasoning Timeout and Process Absence

[git268]

📦 部署环境

Vercel

📌 软件版本

latest

💻 系统环境

Windows

🌐 浏览器

Edge

🐛 问题描述

深度推理模型(如gemini 2.0 flash thingking, o1等模型)在面对复杂问题(数学题)思考较长时间时，不像deepseek-r1那样展示推理过程。并且会因为推理时间较长(大约1分钟以上)而返回错误“没有收到服务器响应”。在openai或gemini官网使用这些模型时能看到推理过程。

Deep reasoning models (such as Gemini 2.0 Flash Thinking, o1, etc) do not show their reasoning process when tackling complex problems (like math questions) and thinking for extended periods, unlike DeepSeek-R1. Often, due to long processing times (around a minute or more), they will return an error such as "No server response received." However, the reasoning process is visible when using these models on the OpenAI or Gemini official websites.

📷 复现步骤

测试的问题为中国2025年研究生入学考试的一道数学题，经过测试gemini 2.0 flash thinking回答这道题需要1-2分钟，o1与deepseek r1需要十分钟以上甚至思考时间超时(哪怕在官网)。

The test question was a math problem from China's 2025 postgraduate entrance exam. In testing, Gemini 2.0 Flash Thinking took 1-2 minutes to answer, while o1 and DeepSeek-R1 required over ten minutes, often timing out even on the official websites.

设 $\Sigma$ 是由直线
\begin{equation}
		\begin{cases}
			x = 0 \\
			y = 0
		\end{cases}
\end{equation}
绕直线
\begin{equation}
\begin{cases}
	x = t \\
	y = t \\
	z = t 
\end{cases}
\end{equation}
(t为参数), 旋转一周得到的曲面,$\Sigma_1$ 是 $\Sigma$ 介于平面 $x+y+z=0$ 与 $x+y+z=1$ 之间部分的外侧,计算曲面积分
\begin{equation}
	\iint_{\Sigma_1} x \, dy \, dz + (y+1) \, dz \, dx + (z+2) \, dx \, dy  .
\end{equation}

🚦 期望结果

希望支持展示更多深度推理模型的推理过程，即使不展示其思考过程也延长从服务器的等待时间。

We hope to see support for displaying the reasoning process of more deep reasoning models. Even if the full thinking process isn't shown, please extend the server timeout period.

📝 补充信息

No response

[lobehubbot]

👀 @git268

Thank you for raising an issue. We will investigate into the matter and get back to you as soon as possible.
Please make sure you have given us as much context as possible.
非常感谢您提交 issue。我们会尽快调查此事，并尽快回复您。 请确保您已经提供了尽可能多的背景信息。

[dosubot]

I found a similar closed issue related to the Gemini 2.0 Flash Thinking model. The issue was resolved by a pull request that added support for the Gemini flash thinking model ^[1].

Regarding your current issue, the lobe-chat framework does have a mechanism for displaying reasoning processes using a tooltip with a specific icon for models with reasoning capabilities. This is indicated by a purple tag with an AtomIcon ^[2]. Additionally, the framework includes a CollapseGroup component that allows for collapsible displays, which could be used for reasoning processes ^[3].

However, I couldn't find specific information on extending server response times for these models or handling long processing times. It might be beneficial to explore the integration settings with external AI model APIs, as the framework uses a streaming approach to handle responses, which can affect response times ^[4][5][6].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}

[arvinxx]

gemini 2.0 flash thinking 在 API中是支持开启 CoT进而流式返回思维链的，但是目前 Google 自己的 sdk 中并不支持开启。可以再等一段时间看看

[lobehubbot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

Gemini 2.0 Flash Thinking supports the opening of COT and then returns the thinking chain, but currently does not support opening in Google's own SDK. You can wait for another time to see

[sxjeru]

https://ai.google.dev/gemini-api/docs/thinking

The Gemini API doesn't return thoughts in the response.
Gemini API 不会在响应中返回 thoughts。

按文档的说法，目前就算换用新的 google-genai SDK，也不会给出思考过程，只能看见空的 thought 字段。

这边在 Python 下试了很多问题，response.thought 一直都是 None 。

[git268]

https://ai.google.dev/gemini-api/docs/thinking

The Gemini API doesn't return thoughts in the response.
Gemini API 不会在响应中返回 thoughts。

按文档的说法，目前就算换用新的 google-genai SDK，也不会给出思考过程，只能看见空的 thought 字段。

这边在 Python 下试了很多问题，response.thought 一直都是 None 。

那么gpt的o1和o3系列呢？我使用微软github或openai提供的o3-mini均不给出思考过程

[sxjeru]

@git268 根据 一些说法，甚至 OpenAI 官网展示的所谓思考过程也是再加工的产物。

从 OpenAI 文档中我没看到任何有关思维链输出的方法，等他们在 API 端发布这个功能吧。

[lobehubbot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

@git268 According to [some saying] (https://x.com/ai_FOR_SUCCCESS/status/1885935217745228014), and even the so -called thinking process displayed by the official website of Openai is also the product of processing.

From the OpenAI document, I did not see any method about the output of the thinking chain, wait for them to release this feature.

[git268]

@git268 根据 一些说法，甚至 OpenAI 官网展示的所谓思考过程也是再加工的产物。

从 OpenAI 文档中我没看到任何有关思维链输出的方法，等他们在 API 端发布这个功能吧。

刚刚发现即使使用deepseek-r1(github提供的671B必定出现，groq提供的70B蒸馏版不会)也会出现类似问题。当给deepseek-r1提供知识库用于检索内容时(图中的知识库为雍正编写的《大义觉迷路》)，deepseek-r1似乎需要思考数分钟才会回答(但该过程不会显示进行深度思考)而显示服务器无响应的超时错误。此外当助手(道教算命排疑解惑系统 https://lobechat.com/discover/assistant/destiny )预设的提示词过长时也会出现同样状况。推测是这种问题不会触发deepseek的深度思考模式但仍然需要较长时间理解提示词或阅读文档。而lobechat似乎20-30秒左右没有收到服务器响应就会中断并显示超时错误

[lobehubbot]

Bot detected the issue body's language is not English, translate it automatically. 👯👭🏻🧑‍🤝‍🧑👫🧑🏿‍🤝‍🧑🏻👩🏾‍🤝‍👨🏿👬🏿

---

@git268 根据一些说法，甚至OpenAI 官网展示的所谓思考过程也是再加工的产物Essence
"
From the OpenAI document, I did not see any method about the output of the thinking chain, wait for them to release this feature on the API side.
"
! [Image] (https://github.com/user-attachments/7e38866a- 421c-9d7821368574)

! [Image] (https://github.com/user-attachments/assets/faa245d3-912a-4279-831b-b8958657458F)))))))
Just found that even if the 671B provided by the DeepSeek-R1 (671B provided by GitHub must appear, the 70B distillation version provided by GROQ will not occur.) Similar problems will occur. When providing a knowledge base for Deepseek-R1 for retrieval content (the knowledge base in the figure is "Lost Lost" written by Yongzheng), DeepSeek-R1 seems to need to think for a few minutes to answer (but this process will not be shown in depth for depth for depth. Think) Show the server without response. In addition, the assistant (Taoist mortal doubt the doubt system https://lobechat.com/discover/assistant/destiny) preset prompts will also occur when they are too long. It is speculated that this problem does not trigger DeepSeek's in -depth thinking mode, but it still needs to understand the prompt words or reading documents for a long time. And the lobechat seems to be interrupted and displayed in about 20-30 seconds without receiving the server response.

[sxjeru]

关于响应超时问题，应该可以通过更换部署方式来解决。

Vercel 限制了 25s 最大等待时间。