重试（稳定）
指数退避加抖动是预期方法；检查当前crate版本以获取暴露的配置API。典型的Rust模式（如果尚未稳定则为伪代码）：
// let policy = RetryPolicy { attempts: 3, base_delay: Duration::from_millis(200), max_delay: Duration::from_secs(2) };
// let client = AiClientBuilder::new(Provider::OpenAI).retry(policy).build()?;
通常重试的错误：网络/瞬态传输问题、HTTP 5xx、带重试头的速率限制。使用幂等键保护副作用工具执行或在模型推理后执行。
重试策略
重试是处理瞬态错误的重要机制，通过指数退避和抖动来避免雪崩效应。
基本配置
use ai_lib::reliability::RetryPolicy;
let policy = RetryPolicy {
attempts: 3,
base_delay: Duration::from_millis(200),
max_delay: Duration::from_secs(2),
jitter: true,
backoff_multiplier: 2.0,
};
let client = AiClientBuilder::new(Provider::OpenAI)
.retry(policy)
.build()?;
高级配置
let policy = RetryPolicy {
attempts: 5,
base_delay: Duration::from_millis(100),
max_delay: Duration::from_secs(10),
jitter: true,
backoff_multiplier: 1.5,
retryable_errors: vec![
"NetworkError",
"TimeoutError",
"RateLimitExceeded",
"ServerError",
],
non_retryable_errors: vec![
"AuthenticationError",
"InvalidRequest",
],
};
使用示例
use ai_lib::{AiClient, Provider, ChatCompletionRequest, Message, Content};
async fn retry_chat(
client: &AiClient,
request: ChatCompletionRequest,
) -> Result<String, Box<dyn std::error::Error>> {
let mut last_error = None;
for attempt in 1..=client.retry_policy().attempts {
match client.chat_completion(request.clone()).await {
Ok(response) => return Ok(response.first_text()?),
Err(error) => {
last_error = Some(error);
if !client.retry_policy().should_retry(&error) {
break;
}
if attempt < client.retry_policy().attempts {
let delay = client.retry_policy().calculate_delay(attempt);
tokio::time::sleep(delay).await;
}
}
}
}
Err(last_error.unwrap_or("All retry attempts failed".into()))
}