Understanding the Harmony Response Format

Understanding the Harmony Response Format

gpt-oss-safeguard uses the [harmony prompt format](https://cookbook.openai.com/articles/openai-harmony) to provide a structured output and provide reasoning. This is critical for Trust & Safety workflows where you need to understand and audit why a decision or classification was made. With the harmony format, oss-safeguard separates its response into two parts:

1. **Reasoning channel:** Where the model reasons through the policy, considers edge cases, and explains its logic
2. **Output channel**: The formatted classification decision you specified

Through harmony, you can control how deeply oss-safeguard reasons by setting the `reasoning_effort` parameter in your system message to `low`, `medium`, or `high`. The model uses `medium` by default if it is not set. Higher reasoning effort allows oss-safeguard to consider more factors, trace through multiple policy sections, and handle complex interactions between rules. Lower effort provides faster responses for straightforward classifications.

If you're using [**vLLM**](https://docs.vllm.ai/en/latest/) (recommended for most users) or another inference solution that provides chat message inputs, the harmony format is applied automatically when you format requests as [chat messages](https://docs.vllm.ai/en/v0.7.0/getting_started/examples/chat.html):

- **System message:** Your policy prompt (include Reasoning: high or similar in the system message to control reasoning effort).
- **User message:** The content to classify.