Skip to content

feat(evals): run the scanned-PDF eval cases with OCR in CI #49

Description

@Fluory

Ziel

The three scanned-PDF eval cases (s01–s03) measure real OCR extraction instead of only the "no text → missing" path. Today the replay gate runs with AI_PDF_OCR=off, because CI does not download the OCR models (#23 / PR #40, #24 / PR #41).

Akzeptanzkriterien

  • CI provides the docling layout and RapidOCR models from an approved source (cached between runs; no download from an unapproved host).
  • The eval replay runs s01–s03 with OCR on and compares against the prepared after_ocr expectations; the baseline is updated once, deliberately.
  • The two docling model tests currently skipped (AI_TEST_DOCLING_MODELS=1) run in CI.
  • The CI step stays within the existing timeouts; a missing model fails the step (fail-closed), it never silently falls back to "OCR off".

Nicht Teil dieser Aufgabe

  • Live model calls (still no credentials in CI).
  • Enabling OCR by default in production.

Betroffene Bereiche

  • .github/workflows/ci.yml (AI service + eval gate steps), services/ai/evals/, services/ai/README.md, PREFETCH_OCR_MODELS

Testplan

Kriterium Prüfart
Models available + cached CI run log
s01–s03 scored with OCR eval gate output in CI
Missing model fails manual CI run with cache cleared / unit test of the preflight

Offene Entscheidungen

Security/Privacy betroffen?

Supply chain: model files from an external host → pin by checksum. Only synthetic eval documents.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    aiLLM, prompts, evalsfeatureNew capabilityreadyDefinition of Ready met – may be claimed

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions