> ## Documentation Index
> Fetch the complete documentation index at: https://docs.verba.ink/llms.txt
> Use this file to discover all available pages before exploring further.

# Response streaming

> Send the beginning of a reply early, then edit it in place as generation continues.

## What the toggle does

On Ultra, turn on **Response Streaming** in **Dashboard -> Verb -> AI Engine**.
On Discord, WhatsApp, and Telegram, Verba sends the first useful part of a
reply as soon as it is available, then edits that same message until the final
text is complete.

This makes long replies feel faster without leaving duplicate partial messages
behind. The completed model response remains the authoritative final message.

## Supported surfaces

| Surface        | Behavior                                                               |
| -------------- | ---------------------------------------------------------------------- |
| Discord        | Sends one initial message, then edits it in place                      |
| WhatsApp       | Sends one initial message, then updates it as supported by the gateway |
| Telegram       | Sends one initial message, then edits it in place                      |
| Verba app chat | Uses its existing live chat/WebSocket response flow                    |
| Public API     | Uses separate SSE behavior described below                             |

## Platform streaming vs API streaming

These are two different features:

* The AI Engine toggle controls platform replies for the connected verb.
* `POST /v1/response` with `stream: true` uses Server-Sent Events (SSE) and is
  currently an Ultra API feature.

Both currently require Ultra, but enabling one does not enable the other.

## Reliability behavior

Verba waits for a useful partial sentence before showing the first update and
throttles later edits so it does not hammer platform APIs. If a provider cannot
stream before anything becomes visible, Verba falls back to the normal
single-message response path.

<Note>
  Streaming improves perceived speed; it does not make the model finish sooner.
  Web search, large context windows, and large outputs can still increase total latency.
</Note>

## When to turn it off

* You only want final, complete messages.
* A platform or proxy is aggressively rate limiting message edits.
* Your verb uses multi-message roleplay where separate messages are intentional.

<Card title="AI engine settings" icon="sliders" href="/guides/ai-engine">
  Configure response streaming, context, reply style, and web search together.
</Card>