Add TextReasoningContent

[stephentoub]

Contributes to #6057

Microsoft Reviewers: Open in CodeFlow

[SteveSandersonMS]

Looks good.

Do you think there should be reasoningtext equivalents for the cases where we concat TextContent:

• ChatMessage.TextReasoning (or maybe better named ChatMessage.ReasoningText as that's what people actually say)
• ChatResponseExtensions.ToChatResponse[Async] coalescing consecutive TextReasoningContent into a single TextReasoningContent

Of these, I'm more interested in the second one, since otherwise streamed output will produce a huge amount of overhead when added to conversation history, having separate content objects for every chunk of reasoning.

[stephentoub]

@tghamm, IIRC, your anthropic library has two custom types for reasoning. Am I correct that this will enable you to replace one or both of them? I think one was for redacted content and the other not, so at a minimum I'd hope this could be used in place of the non-redacted one?

[stephentoub]

Looks good.

Do you think there should be reasoningtext equivalents for the cases where we concat TextContent:

• ChatMessage.TextReasoning (or maybe better named ChatMessage.ReasoningText as that's what people actually say)
• ChatResponseExtensions.ToChatResponse[Async] coalescing consecutive TextReasoningContent into a single TextReasoningContent

Of these, I'm more interested in the second one, since otherwise streamed output will produce a huge amount of overhead when added to conversation history, having separate content objects for every chunk of reasoning.

I will hold off on the first; I'd like us to get more experience with this first. I'll do the second.

[tghamm]

@stephentoub You are correct about the two types, and I THINK using the AdditionalProperties dict would allow for storing the necessary Signature field and perhaps a delineator between Redacted/Normal Thinking Text.

So In short, yes, I believe this would allow for the replacement of at least one, if not both types I had to create.

[RogerBarreto]

Actually the Reasoning Content represents the model Thinking, have we considered using the Thinking name? TextThinkingContent or ThinkingTextContent ?

[stephentoub]

Personally, I prefer "reasoning" over "thinking". These models are generally referred to as reasoning, as this content represents the performed reasoning. @SteveSandersonMS, naming opinion?

[SteveSandersonMS]

Agreed.

Abstractly, "thinking" and "reasoning" are close enough to synonymous that I don't think there's any first-principles reason to say one is more accurate than the other.

Practically, "reasoning" is the word most commonly used for this type of LLM behavior, so I'd side with that. I know there are examples of people referring to "thinking models" and "thinking tokens" but it seems a lot less common than "reasoning models" and "reasoning tokens".

[RogerBarreto]

Thanks for the input, brought this more by experience using the Ollama where they use the first token sharing the Assistant.Content within <think/> xml node.