Problem

Currently, Ballerina AI model providers such as the OpenAI provider lack built-in support for retrying transient failures (e.g., parsing errors) during generate() API invocations.

Proposed Solution

Design document - https://docs.google.com/document/d/1X0Vi93ZCLilkvcYkuhfjx9cGfNKkGMBzo-3edSPB0vA/edit?usp=sharing