Skip to content

feat(eval): NDCG@k with graded relevance #437

Description

@HumanBean17

Tracks deferred work from #431 (fork B). Design ref: docs/superpowers/specs/active/2026-07-12-hybrid-bm25-rrf-design.md.

Why

#431's eval starts with binary relevance (relevant / not). Graded relevance (e.g. exact-symbol match > same-package > same-role) would give a finer ranking-quality signal via NDCG@k.

Scope

  • Add a relevance-grading scheme to the eval ground truth.
  • Implement NDCG@k alongside the existing recall@k / precision@k / MRR.

Why deferred from #431

Binary relevance + MRR is sufficient to evidence #431's thesis; grading adds design + authoring cost better spent once the harness exists.

🤖 Generated with Claude Code

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