Skip to content

RAG Phase 5: 2026 refresh + article-aware chunking (eval-gated) #323

Description

@VforVitorio

Sub of Epic #318. documents/audits/AUDIT_RAG_LAYER.md Phase 5 (size M-L). Eval-gated on Phase 3.

Scope: 2026 refresh + article-aware chunking.

  • Ingest the 2026 FIA PDFs (the ingest side of 2026-reg F-10; coordinate with Epic: Audit 2026-Reg / concept-drift readiness #189, do not duplicate).
  • A/B article-aware chunking to fix the first-regex-match off-by-one article tagging and the wasteful 512-char chunks (bge-m3 accepts 8192), gated on the Phase-3 eval so the change is proven to help.

Deliverable: a 2026-capable, article-correct index whose chunking change beats the baseline on the Phase-3 eval.

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: codebaseTouches src/, scripts/, or notebooks/auditTracks findings from a site/code auditpriority: P2Medium priority — fix this cycle

    Projects

    Status
    Todo

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions