sourcegraph
feedback: please allow editing assistant responses
Version
1.34.0
Areas for Improvement
- [X] UI/UX
- [ ] Onboarding
- [ ] Docs
- [X] Chat
- [ ] Commands
- [ ] Context
- [ ] Response Quality
- [ ] Other
What needs to be improved? Please describe how this affects the user experience and include a screenshot.
When the model replies I sometimes want to nudge following responses into a certain direction / I want to correct mistakes the model made but continue the chat is if those were the model replies. Currently I can only edit my own messages - which will also clear the chat underneath.
Describe the solution you'd like to see
I want to be able to fully control the conversation and edit both user and model messages at any point, then be able to start a new completion at any point. I don't think I can suggest the full UX flow for this but from interacting with LLMs via API it's clear that biasing the model by faking or editing the model responses in a chat session can be extremely powerful (and allow to use fewer tokens than asking it to fix previous messages). This allows better control over the full chat context window.
Describe any alternatives that could be considered
It's also possible to ask the model to change its response to fix things. However this means there are now duplicates in the conversation. The performance also drops as the context gets longer. As a first step being able to delete any message inside the conversation could allow iterating better without blowing up the context window.
Additional context
No response
