Skip to content

Reasoning effort representation #405

Description

@itsaphel

Hey!

Looking at https://github.com/braintrustdata/lingua/blob/main/crates/lingua/src/universal/reasoning.rs, I wonder if the current design (The universal representation uses budget_tokens as the single canonical field.) is optimal

OpenAI uses strings to represent 'effort', Anthropic has deprecated budget_tokens for 4.6 and it's unsupported for 4.7 onwards, and other models (eg Grok, and recent open source like Kimi K3) seem to also use strings. It looks like effort is moving away from being quantified as an integer quantity. I also think the ergonomics of an enum to represent effort is better for a universal representation, to be agnostic of model-specific differences.

I'd probs also note that Anthropic seems to distinguish effort and thinking as separate concepts, but I'm not sure this matters much in practice and whether a universal representation should account for it.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions