Retry

Retry re-runs the wrapped call when it raises a transient error, using a configurable backoff schedule. It is slot-agnostic: attach it as node middleware to retry a whole node, as model middleware to retry a single model call, or both.

By default Retry only retries the transient LLM provider errors (rate limits, timeouts, connection failures). For node-level use, pass your own retry_on tuple.

Usage

import railtracks as rt
from railtracks.prebuilt.middleware import Retry


# Retry is slot-agnostic: use it as node middleware, model middleware, or both.
RetryAgent = rt.agent_node(
    name="retry-demo",
    llm=rt.llm.OpenAILLM("gpt-4o"),
    middleware=[Retry(3)],  # retry the whole node call
    model_middleware=[Retry(3)],  # retry each raw model call
)

Tune the number of attempts, the backoff schedule, and which exceptions to retry:

from railtracks.llm.retries import ExponentialRetry

# Tune the number of attempts, the backoff schedule, and which errors to retry.
picky_retry = Retry(
    approach=ExponentialRetry(max_tries=5),
    retry_on=(TimeoutError, ConnectionError),
)

Ordering

Middleware runs outermost-first in list order. Placed before another middleware, Retry re-invokes everything inside it on each attempt.