deepseek + 搜索代理测试完成,记录下踩过的坑 #110
wolf-joe
started this conversation in
Show and tell
Replies: 2 comments 3 replies
-
当然实际测下来发现模型输出不够稳定,即使是gpt-4o也可能直接给出答案,而不是尝试搜索。 |
Beta Was this translation helpful? Give feedback.
1 reply
-
我也是用的deepseek,有个问题是,一但问题被分解为多个 节点 就会陷入死循环。一直浪费我的tk。 |
Beta Was this translation helpful? Give feedback.
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
我目前用deepseek的api比较多,部署的环境也无法直接访问duckduckgo,部署下来主要踩了两个坑,这里分享一下。
支持deepseek等自定义模型
mindsearch/agent/models.py
,加入自己的模型openai_api_base
这个变量名,字面意义上理解是https://api.deepseek.com/
这种api前缀,但实际要填的是完整的url endpoint。lagent
包(我的路径是./venv/lib/python3.8/site-packages/lagent/llms/openai.py
),让其支持自定义模型不是很理解为啥会限定
model_type
前缀必须是gpt
。使用代理访问duckduckgo
墙内网络环境无法直连duckduckgo,仓库里也没找到传入proxy的地方,所以还是直接改代码(我的路径是
./venv/lib/python3.8/site-packages/lagent/actions/bing_browser.py
):大功告成
Beta Was this translation helpful? Give feedback.
All reactions