Skip to content

Recommendation engine for next action after evals/incidents #37

Description

@glaborie

Summary

Guide users towards highest-impact improvements by suggesting concrete actions after eval, regression failure, or incident review. Examples: change chunk size, add benchmark case, update policy, tune prompt, adjust fallback, tighten tool scope.

Problem to solve

Detection without action is common in ML ops. Users receive pass/fail results but lack expert guidance for improvement. Product value is maximized by bridging the gap from detection to remediation.

Proposed solution

  • On each eval/incident completion, provide targeted suggestions (with rationale and options)
  • Suggest actions on data, prompt, policy, retriever, or model side
  • (Optionally) Integrate with PR regression gate and incident review

Alternatives considered

  • Users research on their own; rely on best practices or vendor consulting
  • External playbooks and workaround docs

Who benefits?

  • Product / operations teams
  • Platform teams
  • AI engineers
  • Maintainers
  • Other

Priority / impact

  • Turns platform from insights to improvement
  • Shortens iteration loops and enhances perceived value

Additional context

Inspiration: Model monitoring products with built-in recommendations, AI-generated pull request suggestions, lint/fix tools in ML pipelines.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions