# AI reply streaming settings

Show AI replies as they are generated and choose what happens to partial text when a reply is interrupted.

*Release 3.3 · admin, developer, operator, visitor*

**Settings › AI › Features**

## What it is

**Enable Streaming** lets a supported AI reply appear while it is being generated. The visible draft is temporary until INSIDE finishes and saves the response. **Interrupt Behavior** determines whether interrupted text becomes a saved partial reply or is discarded.

## Who uses it and why

- Visitors and operators can read text as it arrives on supported chat screens.
- Administrators and developers choose streaming and interruption behaviour for the AI feature.

## Where to find it

Open a supported Autopilot or **Ask AI** feature in **Settings › AI › Features**. Look for **Enable Streaming** and **Interrupt Behavior**. **Team Chat** agents use the streaming setting of their Ask AI feature.

## How it works

1. The AI starts generating a reply and sends available text to the relevant chat screen.
2. The draft grows as more text arrives. Supported content such as images or cards can accompany the response.
3. INSIDE finalises and saves the answer, then replaces the temporary draft with the final message.
4. If the run is interrupted, INSIDE follows **Interrupt Behavior**.

**Persist Partial** saves a cleaned version of the available partial text. **Discard Partial** removes the draft instead. An interruption with no usable text does not create a blank message. A workflow handover discards the buffered reply rather than delivering unfinished text from the previous step.

After a connection interruption, INSIDE can recover an available snapshot of the current reply. Recovery depends on the temporary run still being available.

## Configure it

With the **Settings** role and the **AI** permission, or the **Developer** role, enable **Enable Streaming** on the feature and choose **Interrupt Behavior**. Streaming starts off; the interruption default is **Persist Partial**.

Test both normal completion and interruption on the conversation screen you intend to use. The provider and delivery screen affect whether text arrives progressively.

### Settings

| Setting | Where | Applies to | Default | Who can change it | What it does |
|---|---|---|---|---|---|
| Enable Streaming | Settings › AI › Features | website | Off | Settings role, Developer role | Displays a temporary reply as text arrives, followed by the saved final response. |
| Interrupt Behavior | Settings › AI › Features | website | Persist Partial | Settings role, Developer role | Persist Partial saves cleaned partial text when interrupted; Discard Partial drops the temporary reply. |

### Permissions

| Permission | Lets you |
|---|---|
| Settings | Configure feature streaming when AI is allowed under Settings Permissions. |
| AI | Allows AI configuration within Settings Permissions. |
| Developer | Configure AI feature streaming and interruption behaviour. |

## Limits and good to know

- Some providers return the response all at once, even through a streaming-enabled feature.
- The temporary draft can differ from the cleaned final message.
- Some protected customer-data operations withhold progressive text while access is checked.
- Recovery is temporary. A service restart or expired run can make the draft unavailable; saved conversation messages remain the record of completed delivery.
- Streaming does not expose the AI's private reasoning.

## Depends on and works with

### This feature uses

- **Configured in** [Roles and permissions](ADM-008.md): AI configuration permissions control streaming setup.
- **Configured in** [AI feature settings and change history](AI-001.md): Streaming and interruption policy are selected on each supported feature.
- **Works better with** [Rich content in AI replies](AI-026.md): Final responses can include supported rich content alongside text.
- **Related** [Console connection and recovery](CON-005.md): Reconnecting can recover an available temporary reply draft.

### Used by

- **Required by** [Streaming AI replies for visitors](VX-030.md): The AI reply service supplies streaming, interruption and final response handling.
- **Configures** [OpenAI API connection options](AIP-004.md): The AI feature's streaming options control whether supported incremental output is delivered progressively.
- **Configures** [Google AI Studio model connection](AIP-005.md): The AI feature's streaming options control whether supported incremental output is delivered progressively.
- **Configures** [AWS Bedrock model connection](AIP-006.md): The AI feature's streaming options control whether supported incremental output is delivered progressively.
- **Configures** [Cloudflare AI Gateway connection](AIP-007.md): The AI feature's streaming options control progressive delivery of supported gateway output.
- **Configures** [Autopilot connections through A2A](AIP-019.md): The feature's streaming option controls streaming capability.
- **Enhances** [Autopilot for live chat](AI-003.md): Streaming can show replies as they are generated.
- **Enhances** [Ask AI beside a visitor conversation](AI-008.md): Streaming provides incremental replies and interruption behaviour.
- **Enhances** [AI run limits](AI-014.md): Streaming can display a running reply progressively and control interrupted drafts.
- **Enhances** [Ask AI agents in Team Chat](AI-036.md): Streaming can display the reply while it is generated.
- **Related** [Virtual Assistant tasks started by Autopilot](ENG-027.md): Task handoff discards conflicting buffered or streamed AI reply content.

## FAQ and troubleshooting

### Why does the whole reply appear at once?

The provider may return a complete response, or the current operation may withhold progressive delivery. Check the feature setting and test the configured account.

### Why was unfinished text saved?

**Interrupt Behavior** is probably **Persist Partial**, the default. Choose **Discard Partial** if interrupted drafts should be dropped.

### Why did a draft disappear after handover?

A workflow handover discards buffered output from the previous step. It does not save that unfinished reply.

### Will reconnecting always restore a draft?

No. It can restore available temporary state, but that state can expire or be lost when the service restarts. Check the saved conversation for the final delivered message.