You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Problem
Sometimes I would like to send the contents of the buffer above point sans something irrelevant to the LLM, for example nametags such as "ASSISTANT:" or "USER:" (in a gptel buffer), context that will confuse the LLM such as code not related to the present prompt, or text that should be kept secret.
Describe the solution you'd like
The first thought that jumps to mind is another text property called ignore, which the prompt-making function(s) will skip over.
I don't have any ideas for surfacing this to general users at the moment except perhaps as a gptel-ignore function that applies the property. My thinking is that it would be used with a pre or post response hook to compose the prompt more to the user's liking with a function they write themselves, for example to apply prop 'ignore to text matching some regex.
It may prove useful for other features however, see related below.
Describe alternatives you've considered
To achieve this presently you can create an indirect buffer and manually choose the context you'd like to include in the next message from the original buffer, skirting around what you'd like to avoid sending to the LLM. This method would not work to collect the context as text belonging to its owner, and all of it is sent as though the user wrote it, even if it includes response text.
Problem
Sometimes I would like to send the contents of the buffer above point sans something irrelevant to the LLM, for example nametags such as "ASSISTANT:" or "USER:" (in a gptel buffer), context that will confuse the LLM such as code not related to the present prompt, or text that should be kept secret.
Describe the solution you'd like
The first thought that jumps to mind is another text property called ignore, which the prompt-making function(s) will skip over.
I don't have any ideas for surfacing this to general users at the moment except perhaps as a gptel-ignore function that applies the property. My thinking is that it would be used with a pre or post response hook to compose the prompt more to the user's liking with a function they write themselves, for example to apply prop 'ignore to text matching some regex.
It may prove useful for other features however, see related below.
Describe alternatives you've considered
To achieve this presently you can create an indirect buffer and manually choose the context you'd like to include in the next message from the original buffer, skirting around what you'd like to avoid sending to the LLM. This method would not work to collect the context as text belonging to its owner, and all of it is sent as though the user wrote it, even if it includes response text.
Related: #460, #408.
The text was updated successfully, but these errors were encountered: