Retries

When a model call fails, chat-rs can retry it. You set how many attempts to make and, optionally, a retry strategy that runs between attempts so you can back off, log, or react to the specific failure.

Set it up

use chat_rs::{ChatBuilder, openai::OpenAIBuilder};

let client = OpenAIBuilder::new().with_model("gpt-4o").build();

let chat = ChatBuilder::new()
    .with_model(client)
    .with_max_retries(5)
    .with_retry_strategy(chat_rs::retry_strategy!(|ctx| {
        println!("retrying, attempt {}", ctx.idx);
        tokio::time::sleep(
            tokio::time::Duration::from_millis((400 * ctx.idx).into()),
        )
        .await;
    }))
    .build();

The retry_strategy! macro builds the closure for you and exposes ctx, a CallbackRetryContext:

pub struct CallbackRetryContext {
    pub idx: u16,            // attempt number, starting at 0
    pub failure: ChatFailure, // the error that triggered this retry
}

The strategy runs after a failed attempt and before the next one, so it is the right place for exponential backoff, jitter, or rate-limit handling. Inspect ctx.failure to decide how long to wait, or to log what went wrong.

How it relates to max steps

with_max_retries controls how many times a single model call is retried after an error. It is different from with_max_steps, which controls how many times the agentic loop goes around while the model is calling tools. One handles failures; the other handles tool-driven turns.