Kagi Assistant "sees" earlier reply after you edit your message.

Value7609

This can be easily checked, by first asking a question, then editing that question to say "What did you say in your last message?"

I expected it to only "see" the messages above my question.

I'm often editing the message, because I don't like the reply it gave me, but it gets very repetitive, as the LLM still "sees" the replies it gave to my question earlier which I edited.

Luis

Value7609 to clarify, you're reporting this as a bug because you're editing the first message in the Assistant thread, and since it's the first message, editing it should work similarly to creating a new thread, correct?

Value7609

Luis yes in this case, but lets say I already sent 3 messages and got 3 answers, now I'm editing my third message, and the assistant should only "see" my first 2 messages, its 2 answers + my third message, but definitely not the third answer.

Thibaultmol

interesting, I can't repro it with gpt 4o but I can with sonnet 3.5

cardinal086

When clicking edit to branch of the conversation all messages are in the context even the ones after the "edited" prompt.

To replicate do something like "generate 10 random words".
Then edit the prompt and ask "what was the last random word?"

It will correctly respond with the last word even if the generation of words was deeper in the conversation.

The context should only consist of messages before the edited prompt.

Luis

cardinal086 Hey! Merging this conversation with a similar thread to streamline our communication

fxgn

This is incredibly annoying, because it also happens when clicking the "regenerate" button. I have to create a new chat with the same prompt every time Assistant says something wrong, because I don't want it to see it's previous incorrect response.

IsaacThoman

fxgn YES! It doubles down on its own previous nonsense!!

Value7609

Luis I don't know if this is related to the same issue. For me these are two different things.

This one is about it seeing a reply which does not exist in this "timeline".

The other one you've merged is more of a visual issue, as with pressing regenerate my previous prompt stays the same, that's why my prompt shouldn't be branched, but the LLMs response should be branched. Otherwise it looks like I have edited my message (because I can see the branch buttons on my message) contrary to the fact, that my message stayed the same, but the LLM ""edited"" its message (as in regenerated).

Luis

Value7609 Oh, my bad—I misunderstood your post. It seemed odd you'd create the same issue twice :p Sorry about that, I'll split it.

IsaacThoman

The chat assistant still has context of my previous message after I edited the message. What's the point in having an edit button if it's just appending my edit to the bottom of the chat? This is a bug.

The new message should replace the old one in the context, not be appended to the bottom of the context.

Luis

IsaacThoman merging with a similar report to streamline communication