Swipes: regenerating alternative replies

A swipe is an alternative version of the latest AI reply, generated on request, that you can flip between to pick the one you want to keep.

What swipes are

In SillyTavern and similar frontends, the last AI message has arrows on its sides. Swiping right generates a new reply to the same prompt and stores it alongside the old ones; swiping left goes back. Whichever swipe is showing when you send your next message is the one that becomes part of the history. The other versions stay saved in the chat file but are not sent to the model.

Alternate greetings on a character card appear as swipes on the first message.

How a swipe works

A swipe is a fresh request with the same prompt as the reply it replaces. Because sampling is random, each attempt takes a different path. How different depends on your settings: at low temperature or with tight truncation, swipes look nearly identical; with higher temperature and a sampler like min P, they vary more.

What swipes cost

Every swipe resends the full prompt. In a long chat that might be tens or hundreds of thousands of input tokens per swipe, plus the new output. Five swipes on a message cost about five times one reply. Two things reduce that:

Prompt caching, where supported, because a swipe repeats exactly the same prompt.
Stopping a bad reply early while streaming, so you do not pay for output you will throw away.

Using swipes well

If several swipes in a row make the same mistake, the problem is in the prompt, not luck. Fix the card, author's note or your last message instead of swiping more.
Edit a nearly good swipe rather than rolling dice for a perfect one. Edits become history the model will imitate.
Use swipes as a sampler test: run five at the same settings to see the spread before changing values.
Swipe sooner rather than later. Once a weak reply is followed by your answer, it becomes history, and editing or deleting it is the only way back.
Be aware that picking only the most extreme swipes trains the chat's style toward extremes, since later replies copy earlier ones.

Swipes with Wild West API

Each swipe is a normal billed request to /v1/chat/completions. Some APIs support an n parameter to return several completions in one call, but frontends usually generate swipes one at a time, and support for n varies by upstream, so do not rely on it.

FAQ

Why are my swipes all the same?

Your sampling is too deterministic. Raise temperature a little or loosen top P or min P, and check that no seed is fixed.

Do swipes cost money?

Yes. Each swipe is a full new request, so you pay for the whole prompt again plus the new reply.

Temperature: how it changes what the model writesTemperature is a sampling setting that makes a language model's word choices more predictable when lowered and more varied when raised. Prompt caching: reusing the start of a promptPrompt caching is a feature where a model server saves the computed state of a prompt's opening section so later requests that start with the same text can skip that work. First message: the greeting that sets the toneThe first message, or greeting, is the opening reply from the character that starts every new chat and sets the style the model will imitate. Streaming responses from an LLM APIStreaming is an API mode where the model's reply is sent piece by piece while it is being generated, instead of all at once at the end.