S2 pilot handover¶
AMČR colleagues run the six ATRIUM services themselves in the S2 pilot (December 2026 – January 2027). This page is where the handover starts: which images, which runbook for each, what to watch for, and whom to ask. It follows the handover rules of the acceptance runbook.
The six services¶
The versions below are each tool's newest release on 25 October 2026. 11 November is the target for
choosing and approving the six -api images. AMČR approves them explicitly, by digest, and the digests
are entered here once approved; until then the column stays empty, and latest or test are never the
baseline. Acceptance on AMČR's cluster follows separately (below).
| Stage | Image | Release | Approved digest | Runbook | Issues |
|---|---|---|---|---|---|
| page categories | ghcr.io/ufal/atrium-page-classification-api |
1.9.10-beta |
— | service/README.md | tracker |
| OCR output and text files → lines | ghcr.io/ufal/atrium-ocr-postprocess-api |
1.12.1-beta |
— | service/README.md at f7341b4 (on test; released with the next tag) |
tracker |
| born-digital files → the record | ghcr.io/ufal/atrium-digital-convert-api |
1.2.2-beta |
— | service/README.md | tracker |
| morphology, named entities, TEITOK | ghcr.io/ufal/atrium-nlp-enrich-api |
1.1.3-beta |
— | service/README.md | tracker |
| keywords | ghcr.io/ufal/atrium-keyword-extract-api |
1.4.0-beta |
— | service/README.md | tracker |
| English metadata | ghcr.io/ufal/atrium-translator-api |
1.5.0-beta |
— | service/README.md at 1066070 (on test; released with the next tag) |
tracker |
digital-convert 1.2.2-beta is the first release that certifies its images: its images.json asset names
the digest of each image it published (#81); the other five do so from their next releases. The image tag is the release without its leading v. A digest is read with
docker buildx imagetools inspect <image>:<tag>. Every E2E run also records the digest and the source
ref of each image it ran (e2e-manifest.json, an artifact of the run): a green run at these release
tags shows which releases passed together. The lanes run the batch images for most stages, so the
approved digests are still those of the -api images of the same releases.
Before you rely on a result¶
Each runbook ends with a Known limits section. The ones a pilot user meets first:
- ocr-postprocess scores language and quality line by line; its rules are tuned for Czech, so lines in other languages are scored less reliably.
- translator refuses a language pair its backend has no model for (422,
cause: unsupported_language_pair) rather than routing it through another language; by default it calls LINDAT, whose output terms are open under #78. - digital-convert reads born-digital files only: a PDF that carries an OCR text layer is refused, and pages whose text was extracted damaged are flagged for OCR. Its upload limit is 250 MB; a deployment can raise it for larger originals, but files up to 1.5 GB are not yet tested for memory, concurrency and timeouts.
- keyword-extract's vocabulary-controlled kind is experimental and not evaluated; its statistical kind needs no backend.
- page-classification classifies at most
MAX_PDF_PAGESpages of a PDF per request (default 50); set it to the longest report when it runs behind digital-convert's/describe. - nlp-enrich calls LINDAT's UDPipe and NameTag. Its
/jobsare kept in memory: a restart loses them, and the caller submits a lost job again (accepted for the S2 pilot).
Acceptance¶
AMČR's acceptance on its own cluster is planned for the second half of November, before the December pilot, coordinated with @rharasim. The procedure and its results table are in the Kubernetes acceptance runbook: one row per service, with the digest each pod ran and what each check printed, and for the translator a run on real AMČR metadata. The completed table is linked here when AMČR returns it.
Contacts¶
| Role | Who |
|---|---|
| the six tools, their releases and runbooks | @K4TEL |
| LINDAT services; ÚFAL repository settings | @stranak |
| AMČR pilot lead | @motyc |
| AMČR acceptance on the cluster | @rharasim |
Further AMČR contacts are added as AMČR names them. Questions go to the issue tracker of the service concerned, or to #77 for the handover as a whole.