Area(s)

area:gen-ai

What's missing?

Source semconv: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-events/#event-gen_aichoice

Since the release of the DeepSeek R1 model, an increasing number of model providers have started offering reasoner-type LLMs. Some of our users are paying attention to certain new elements emerging in LLM invocations, such as reasoning content.

Of course, for different vendors, there are still many different detailed context. For instance, whether reasoning tokens should be counted separately, or whether the "reasoning effort" attribute should be recorded. I believe it may not be the right time to discuss these topics yet. However, when it comes to reasoning content, the answer is clear — this parameter is highly valuable for aiding in evaluations related to chain-of-thought reasoning.

Reference:

1. https://api-docs.deepseek.com/guides/reasoning_model
2. https://platform.openai.com/docs/guides/reasoning
3. https://docs.dify.ai/learn-more/use-cases/integrate-deepseek-to-build-an-ai-app

Describe the solution you'd like

Add the capture of message.reasoning_content in gen_ai.choice in case of more and more llm sdks/orchestration frameworks are supporting for this field.