Anonymous 01/19/25(Sun)12:58:43 | 3 comments | 🔒 Locked
Are there stuff for the Open WebUI that take user input, check with one model to determine the input's context, and then select the best model for a response? For example, if the initial prompt sets a roleplay character or specific helper role, and you ask in Chinese for a Bash script that translates English to Chinese and generates an image about the basic logic of the script, the system could:
>use a roleplay model to maintain the character
>select the best model for Chinese translation
>use a coding model for the script
>generate the image with the best model
>reflect and refine the output using another model in 1–10 iterations
It could also use a model to critique and refine responses to avoid looping or breaking character.
>use a roleplay model to maintain the character
>select the best model for Chinese translation
>use a coding model for the script
>generate the image with the best model
>reflect and refine the output using another model in 1–10 iterations
It could also use a model to critique and refine responses to avoid looping or breaking character.
Anonymous 01/19/25(Sun)13:05:48 No.103957213
Also why there isn't any simple fixes for the main problems in the basic AI features? For example the web search is shit (yes, even your favorite spyware like latest ChatGPT has it):
>user input asks something and gives an url
>ai says "I'm unable to access the specific thread at the provided URL" because website has limitations for automatic scripts to harvest the data
Why there isn't some simple docker container that browses websites like a normal human with graphical GUI and uses user's cookies for this?
>user input asks something and gives an url
>ai says "I'm unable to access the specific thread at the provided URL" because website has limitations for automatic scripts to harvest the data
Why there isn't some simple docker container that browses websites like a normal human with graphical GUI and uses user's cookies for this?
Anonymous 01/19/25(Sun)13:09:10 No.103957259
Anonymous 01/19/25(Sun)13:12:52 No.103957307
>>103957259
I was more interested why there isn't one already. There must be some reasons why all of this is a shit idea in general. It would seem so basic to assume that best way would be to do every step with best model that can run with 24 GB of ram. Sounds stupid to make models like Deepseek v3 that needs 1,5 TB of VRAM to run and can't even do something basic as speak count 2+2 when asked to do so in Finnish or can't read web url when user gives one.
I was more interested why there isn't one already. There must be some reasons why all of this is a shit idea in general. It would seem so basic to assume that best way would be to do every step with best model that can run with 24 GB of ram. Sounds stupid to make models like Deepseek v3 that needs 1,5 TB of VRAM to run and can't even do something basic as speak count 2+2 when asked to do so in Finnish or can't read web url when user gives one.