[Feature Request] Enhance conversation repl

[sigoden]

GPT-3.5 supports up to 4096 tokens, and it is not very likely to have a long conversatio, so there is no need to design a feature-rich conversation repl.

But gpt-4 supports up to 8k tokens, gpt-4-32k supports up to 32k tokens, we can do long conversation. It's necesarry to enhance conversation repl.

.conversation - show conversation help

.conversation new - create a new conversation

.conversation revert [num] - revert current conversation messages.

This give user the ability to edit previous lines, thus branching the conversation in a new direction. This is useful when models get caught in infinite loops, or local optima.

The num parameter specify how many messages to revert. If we omit num parameter, Aichat will display a list of current messages, the user can choose where to revert.

.conversation delete - delete a saved conversation

AIChat will display a multi-select list for the user to select the conversation to delete

.conversation use - reuse a saved conversation

AIChat will display a list for the user to select the conversation to load to

Design details

1. The conversation automaticlly saved to $CONFIG_DIR/conversations/<datetime>-<first-message-truncated>.json
2. The conversation is not bound to a specifiy model.
3. Allow to run .model in the middle of conversation.
4. Not allow to run .role or .prompt in the middle of conversation.

Welcome to provide suggestions for the design of the conversation repl.

[crabdancing]

.conversation is a bit long to type every time. Shortening it to .conv would be helpful -- perhaps as an optional abbreviation?

[sigoden]

aichat's readline supports tab completion, type .conv then type tab, you get .conversation. there is no need to use abbreviation.

@alxpettit

[asb]

One related request/suggestion I have would be to make it easy to convert a one-off request into a conversation. Often I expect a one-off response to be fine, but upon getting it back I realise I have a follow-up question or refinement, and this may be better handled through a conversation rather than editing the prompt.

As a strawman, how about .conversation last to start a conversation that incorporates the last question and answer.