Triage-to-Arbitration LLM Pipeline With OpenAI Models

OpenAI's official reasoning best practices document a two-stage e-commerce workflow: use fast GPT-4o and GPT-4o-mini models to triage order details with customer information and identify order issues and return policy clauses, then feed all of those data points into o3-mini to make the final decision about return viability.

Stage 1 (GPT-4o-mini): Triage order details, customer information, identified order issues, and matching return policy clauses into <triage_data> tags. Stage 2 (o3-mini): Given the extracted <triage_data> points, evaluate all policy conditions and output the final binding decision on the viability of the return.