🚀 Feature: Add thinking tokens to the instrumentation

[dinmukhamedm]

Which component is this feature for?

Anthropic Instrumentation

🔖 Feature description

We should track the content of thinking tokens for thinking models that return those, such as claude-sonnet-3-7-thinking

🎤 Why is this feature needed ?

It is extremely helpful in debugging prompts, especially for larger and more complex tasks

✌️ How do you aim to achieve this?

Anthropic returns its "thoughts" in a separate field. The only unclear thing is how to name the attributes.

I propose storing the thought as a separate role, i.e.

• gen_ai.completions.0.role = "thinking"
• gen_ai.completions.0.content = "I am asked to ... I should ..."

🔄️ Additional Information

No response

👀 Have you spent some time to check if this feature request has been raised before?

• I checked and didn't find similar issue

Are you willing to submit PR?

Yes I am willing to submit a PR!

[nirga]

Thanks! @dinmukhamedm! We definitely need this :)
(Wondering if OpenAI has something similar)

[dinmukhamedm]

Awesome! I'll open a PR later this week. I think OpenAI doesn't return it's thinking contents by design (as of now)

[nirga]

I think they do now with the new response api (for which we also need to add support)

[dinmukhamedm]

@nirga I created #2782 for this