Stream AI responses without treating partial text as a final answer
Handle documented stream events, show an unfinished state and wait for completion before parsing or acting on the result.
In brief
Streaming sends output while it is generated. It can shorten the wait before a reader sees something, but it does not establish that the full answer arrived or reduce every part of processing time.
A clear unfinished state prevents a network interruption from looking like a complete concise answer.
Show progress honestly
Render text fragments safely and retain an in-progress indicator. Avoid injecting generated HTML into the page without appropriate sanitization. Provide a way to cancel where supported, and do not promise that cancellation means no usage was billed.
Respect event boundaries
Follow the documentation for the selected API rather than concatenating every event body as visible text. Keep tool arguments separate from prose and validate the completed object. Do not send a partial draft email merely because the first words look acceptable.
Handle incomplete output
If the network fails, label the answer incomplete and offer an appropriate retry. Keep enough non-sensitive state to avoid duplicate actions. For structured extraction, wait for final completion and validate before updating a record.
Before you act
Different stream events can contain text, tool calls, errors or completion information. A fragment of JSON or a tool argument must not be executed as though it were final.
What to do next
- Use the API’s supported streaming route and event types.
- Accumulate partial output while displaying that generation is in progress.
- Handle final success, error, refusal and interruption explicitly before storing or acting on the result.
