Retry wrapper for LLM calls with exponential backoff and jitter.
Retries on transient errors (HTTP 429, 500, 502, 503, 504, connection errors). Does NOT retry on client errors (400, 401, 403, 404).
Options
:max_retries- Maximum number of retry attempts (default: 3):base_delay_ms- Base delay in milliseconds (default: 1000):max_delay_ms- Maximum delay in milliseconds (default: 30_000)
Examples
ADK.LLM.Retry.with_retry(fn -> ADK.LLM.Gemini.generate(model, request) end)
ADK.LLM.Retry.with_retry(
fn -> ADK.LLM.Gemini.generate(model, request) end,
max_retries: 5, base_delay_ms: 500
)Summary
Functions
Compute backoff delay with jitter for the given attempt.
Extract a retry-after delay (in milliseconds) from a Req.Response.
Returns true if the error is transient and should be retried.
Execute fun with retry logic. Returns the first success or the last error.
Functions
@spec compute_delay(non_neg_integer(), pos_integer(), pos_integer()) :: non_neg_integer()
Compute backoff delay with jitter for the given attempt.
@spec extract_retry_after(%{headers: term()}) :: non_neg_integer() | nil
Extract a retry-after delay (in milliseconds) from a Req.Response.
Checks for Anthropic's retry-after-ms header first, then the standard
retry-after header (interpreted as seconds). Returns nil if neither
is present or parseable.
Returns true if the error is transient and should be retried.
@spec with_retry( (-> {:ok, term()} | {:error, term()}), keyword() ) :: {:ok, term()} | {:error, term()}
Execute fun with retry logic. Returns the first success or the last error.
fun should return {:ok, result} or {:error, reason}.
Options
:retry_after_ms- If set, overrides the computed backoff for the first retry attempt (useful when the server sends aretry-afterheader).:max_retries,:base_delay_ms,:max_delay_ms- see module docs.:sleep_fn- override for testing (default:&Process.sleep/1).