Sora has developed Deep Refine, a Python tool designed to enhance AI Large Language Model (LLM) responses through strategic, multi-model refinement cycles. When given a prompt, Deep Refine processes it through a configurable sequence of language models, each utilizing carefully crafted prompting strategies. The system initiates by sending an initial prompt template—containing the user's request—to the first LLM. Each subsequent model receives both the original user request and the previous model’s response, wrapped in a specialized refinement prompt template that guides the model to critically evaluate and improve upon the existing answer. This cycle continues through all selected models in sequence, with one complete loop through all models constituting a single refinement iteration. Deep Refine will proceed through as many refinement iterations as the user specifies, or until response convergence is detected.
The tool supports major API providers, including OpenAI, Anthropic, Google, Chutes.ai, and OpenRouter.
Deep Refine also generates a comparative analysis of how the response evolves across refinement iterations. Results are organized in a directory structure that preserves the complete refinement history. For complex queries, the generated refinement history is particularly useful for observing the unique strengths of different Large Language Models as they work in sequence to produce a result superior to what any single model could achieve alone (by default).
The end result is a Python tool that progressively refines language model outputs through multiple iterations. Deep Refine takes a complex question, sends it through a sequence of AI models, and then feeds each response back with instructions to critically analyze and improve the previous answer. This process continues until the configured iteration limit is reached or convergence is detected when further improvements become minimal.