Skip to content

perf: decrease cost per highlight #128

@MartinBernstorff

Description

@MartinBernstorff

Instrumentation:

  • Track token-use and log it

Experiments:

  • Skip answer-generation

  • Generate questions for all new highlights in a document at once – avoids:

    • Independent reasoning for each highlight
    • Duplicate Q/A pairs
  • Generate reasoning for all highlights in a document at once, and re-use across highlights

  • Frame the reasoning more tightly, to use fewer tokens. E.g. must be bullet-points, 5 headers, no summary.

  • Integrate the reasoning and question generation steps

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions