Recover from transient startup failures
Transient provider or network failures occur before a model call starts.
AI Tools · Const
Retries
maxRetries controls how many times the AI SDK retries failures while starting a model call. It defaults to 2 and can be set to 0 to disable retries.
How it works
Retries are configured through the request option maxRetries, which limits retries for failures encountered while starting a model call. Set it to zero to disable them; stream recovery after streaming begins is a separate feature controlled by streamRetries.
Syntax
maxRetries?: numberParameters
| Name | Type | Description |
|---|---|---|
maxRetries | number | Maximum number of retries. The default is 2; use 0 to disable retries. |
Key points
maxRetriessets the maximum number of retry attempts for a request.- It covers failures during model-call startup.
- Use
streamRetriesseparately for recoverable provider errors detected after streaming begins. - This setting belongs to request/transport configuration rather than generation parameters.
Common mistakes
- Retrying can increase response time and provider charges.
- Use zero when you need a single attempt.
maxRetriesdoes not enable recovery for provider errors that arrive after streaming starts.
Example: Configure model-call retries
const result = await generateText({
model: 'openai/gpt-4o',
prompt: 'Write a haiku about retries.',
maxRetries: 3,
});This requests up to three retries when starting the model call.
When to use it
Disable retries
A request must fail immediately without another attempt.
Configure stream recovery separately
A streaming response can emit a provider error after streaming begins.
Related
streamRetriesRequestOptionsgenerateTextonError