[Question] How to track/log the current retry attempt when using request retries?

[leoplct]

Hi there, thanks for the great gem!

I am using openai (v0.42.0) with Rails 8. I have configured the client to handle retries automatically, but I am struggling to find a way to "observe" the retry process in real-time.

The Context I am running a background job that extracts data from invoices. I want to be able to log (or update a database record) every time a retry occurs (e.g., "Attempt 2/5 failed, retrying..."), specifically when hitting 429 Rate Limits.

Currently, the .chat or .completions call is blocking, and I only get the result after all retries are exhausted or successful. I have no visibility into how many attempts were made during the execution.

client = OpenAI::Client.new(
  access_token: "...",
  request: { timeout: 20 } 
  # I understand retries are handled via Faraday middleware under the hood
)

# How can I hook into the retry mechanism here?
@response ||= $openai.responses.create(
      model: MODEL,
      service_tier: service_tier,
      store: false,
      instructions: "Extract details from the provided document.",
      reasoning: { effort: "medium" },
      input: [
        { "role": "user", "content": inputs }
      ],
      text: {
        format: {
          "type": "json_schema",
          "name": "extraction",
          "schema": Openai::Schema.schema(@input)
        }
      }
    )

The Question Is there a supported way to pass an on_retry callback or a block to the underlying Faraday connection through the OpenAI::Client?

I would like to achieve something similar to Faraday's retry block:

retry_options = {
  max: 5,
  retry_block: -> (env, options, retries, exception) { 
    puts "Retrying... Attempt #{retries}" 
    # MyCustomLogger.log(retries)
  }
}

Is there a configuration option I missed to expose this?

Thanks for your help!

[ms-jpq]

hi @leoplct,

We have plans to implement middlewares for the SDK such that you will be able to do this easily. In the interim, you would unfortunately have to subclass the OpenAI::Client

class MyClient < OpenAI::Client

   def request(...) = ...
end