Repository navigation
Releases
DM adaptation to the task + docs refined
Minor release on top of v1.4.5-beta. Conversions, the service and the record are unchanged. What is new is offline tooling for the document-type decision models (#8, atrium-project#83); neither the service nor the images run it. v1.4.6-beta carries the same tooling but was never released: on 2026-11-08 GitHub's release downloads answered 504, and the release workflow could not install Trivy.
Added
tools/build_document_types.pywritesdecisions/document_type.v2.json: AMČR'sdokument_typ(32 of its 41 terms in 6 groups, withneurčený dokumentas the abstain option). It has one question set per input:pdf_page: the rendered page image, with its text layer only on request;document_text: keyword-extract's view, OCR text only, so it offers the 6 written-document types and the abstain option.
tools/decide_documents.pyasks a decision model atPOST /v1/systemonefor the type of each PDF page and of each document (from at most its first 4 pages).- It reads the image alone by default;
--with-textadds the text layer. - It writes
DECIDE_PAGES.csvandDECIDE_DOCS.csv, with a summary and paradata. --goldand--page-goldscore the answers.
- It reads the image alone by default;
decision_client.py, vendored from page-classification v1.13.0-beta and pinned bytests/test_vendored_decision_parity.py.decision_modelinpara_config.txt(Apache-2.0, conditional).DECISION_MODEL_URL,DECISION_MODEL_NAMEandDECISION_TIMEOUT_Sare defaults of the offline CLI only, not service settings.- Documentation: both tools in
docs/usage.mdanddocs/development.md.
Fixed
decide_documents.pynow exits 2 before any request when:--accept-ator--review-belowis outside 0–1;--review-belowis above--accept-at, which used to fail at the first answer, after a model call;--doc-pagesis outside 1–4, which used to be clamped while the paradata recorded the value given.
Changed
- A manual run of
release.ymlon a tag now creates the release; this is the recovery when a tag push starts no run. On a branch it is still a dry run. Tagging uses the hub'sscripts/tag_release.py. - CI: the hub's workflows are pinned to
4435ffb(Trivy from its ghcr.io image; the OpenAPI baseline through the GitHub API).atrium_openapi.pyis re-vendored from that commit. - Dependencies: python-multipart ≥ 0.0.33 in
service/requirements.txt. data_samples/README.mdnow names the run the samples come from (2026-10-18, v1.2.0-beta).
Compatibility
- No service response, no record field and no part of
service/openapi.jsonother than its version changed. - keyword-extract pins
decisions/document_type.v2.jsonat commit427b690; this is the first release that carries it.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.4.7-beta=sha256:04a89e8bee221c4edc1abd075ed5db4b99198cfc305ea98daec11d15d5a987aaghcr.io/ufal/atrium-digital-convert-api:1.4.7-beta=sha256:6bec42f1ce48cea7e0b46e07079a13e38e0649ceaa63b087e1db5e8307203f25
Full Changelog: v1.4.5-beta...v1.4.7-beta
Preparation for Galaxy wrapper integration
Patch release on top of v1.4.4-beta. Conversions are unchanged.
Changed
- The command-line image
ghcr.io/ufal/atrium-digital-convert-digital, and the locally builtdigital-doclingstage, start through the hub-canonicalcontainer_entrypoint.py, which runs/app/api_util/digital_to_json.pyby its absolute path (atrium-project#66). Arguments still go to the converter; a first argument that namessh,bash,dash,pythonorpython3runs as given, which is how Galaxy starts a tool's container (docker run IMAGE /bin/sh tool_script.sh). Before, Galaxy's/bin/shbecame the input file (exit 2), and the relativepython api_util/digital_to_json.pyfailed outside/app. - Only those interpreters run, so a document that happens to be executable is still converted.
- Comments in
requirements_digital.txtanddigital_born/probe_pdf_geometry.pysayprovenance.licenseis a calculated label, not the output's licence (atrium-project#6). - Tests:
tests/test_batch_entrypoint.pypins thedigitalanddigital-doclingstages;tests/test_container_entrypoint_shared.py, vendored from the hub, tests the module. - CI: the hub's workflows are pinned to
da7a066, the commit that addscontainer_entrypoint.py.
Compatibility
- The default
--helpand every conversion command behave as before. Only a first argument that names one of the five interpreters changes meaning: it now runs that command. The-apiimage is unchanged.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.4.5-beta=sha256:fa7fc0366074e070463e4aab075edcfc079d9e57407e143f4c98d228d0f82b84ghcr.io/ufal/atrium-digital-convert-api:1.4.5-beta=sha256:891699096e8c2abbc1c47a4ccf56cb6ddc1352e7e2f4160aada09b58e12a96e1
Full Changelog: v1.4.4-beta...v1.4.5-beta
RO-Crate and GHA fixes
Patch release on top of v1.4.2-beta. v1.4.3-beta was tagged but not released (its tag push started no workflow run); this release carries its changes. Conversions are unchanged.
Changed
- An RO-Crate export no longer presents the calculated licence as the crate's: the root's
licensenames#no-licence-asserted, an entity that asserts nothing (the rights depend on the source and on AMČR's permission). A fragment carries no licence entity (atrium-project#6). - A record marks its calculated licence with
provenance.license_basis: "calculated", declared in the record schema and inservice/openapi.json. - The paradata's licence notes list the components and say that no output licence is asserted; they no longer read "Effective output license is X, the most restrictive among N component(s) used".
- CI: the hub's workflows are pinned to the commit whose Trivy scan and CVE gate each run once more after a 30 s pause (atrium-project#81).
Compatibility
provenance.license,license_urlandlicense_detailare still written, and records stay valid underdoc-schema-v1. A consumer that read a crate root'slicenseas a licence now gets#no-licence-asserted.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.4.4-beta=sha256:839a2263f0c613dc73cc55d4fd941779712bc00007ca653afd65b0038e32b115ghcr.io/ufal/atrium-digital-convert-api:1.4.4-beta=sha256:57ebac40ee78dcc5dbe7e0b8d862a23d81f6f2d91c3ad3994980c829b0c38cdb
Full Changelog: v1.4.2-beta...v1.4.4-beta
Documentation refined + GHA edits + licensing changed
Patch release on top of v1.4.1-beta.
Fixed
- uvicorn follows
LOG_LEVEL: underWARNINGit still logged its start-up lines and every request.
Changed
- flexiconv is pinned to v0.3.11, its first tag under the MIT licence (re-vendored from nlp-enrich v1.3.2-beta); conversions are unchanged.
para_config.txtdeclares it MIT (atrium-project#6). - The documentation lists a record's components instead of a calculated licence (atrium-project#6) and is arranged by reader (atrium-project#79).
- CI: the hub's workflows are pinned to
6a10787. The weekly fresh-environment lane installs Tesseract and the OCR-fallback requirements, and fails when flexiconv does not install.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.4.2-beta=sha256:be4e4a5612a2cf902a7c8d1c1d87ddab451022550c6a82bc273157d6558c3ab9ghcr.io/ufal/atrium-digital-convert-api:1.4.2-beta=sha256:4da95a351576195ae486c4fae0b45de5162e9a40c0cf642dbf04f14e27de667e
Full Changelog: v1.4.1-beta...v1.4.2-beta
GHA fixes + shared scripts update
Patch release on top of v1.4.0-beta, and the S2 pilot candidate (atrium-project#77).
Fixed, on both entry points
- A
document_json/--document-jsonrecord whosepagesorlinesis not a list of objects is refused: 422invalid_recordfrom/reformatand/describe, exit 3invalid_recordfrom the command line. It used to end in a 500 and a traceback (atrium-project#72). - v1.4.0-beta's notes say an inherited record that does not validate is added to with a warning. That held except in this case, which crashed.
Changed
- The
invalid_recorddescription inopenapi.json, the exit codes in--helpanddocs/usage.md, and the parity table name this case. - CI: the hub's workflows are pinned to the commit with the shared fix, including the Trivy database retry.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.4.1-beta=sha256:d92ee2164fd784aa9c9bc34233ff30ab98f7388e66d6ce23b8f58fd6db923b0eghcr.io/ufal/atrium-digital-convert-api:1.4.1-beta=sha256:e7f295de0b228ca7e995026f3ca13860a518ff6522c6e555796336c489ad779a
Full Changelog: v1.4.0-beta...v1.4.1-beta
API-CLI alignment of results
v1.4.0-beta: the command line and the service end alike
What changed (atrium-project#72)
- An empty file is
corrupton both sides; the command line saidunsupported. - A
--document-jsonrecord that is not JSON, or has a newer schema major, isinvalid_record(exit 3), as the service's 422. Before, the command line ended in a traceback. - An inherited record that does not validate is added to with a warning, as the service answers 200 with
document_json_schema_error. Before, the command line crashed. - One output check and one paradata builder serve both sides, so the paradata's parameters and output counts match.
docs/usage.mdhas one table of what still differs between the two entry points, and why;--helplists the exit codes.
Compatibility
- The API is unchanged apart from the version. Command-line exit codes: an empty file 4 (was 3); an unreadable seed record 3 (was 1).
Upgrade notes
- Pin
1.4.0-beta.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.4.0-beta=sha256:74fe672f37b6fb8442d3bcfa293eec65e0d61d7f3d0ba3f03867692f97c154a6ghcr.io/ufal/atrium-digital-convert-api:1.4.0-beta=sha256:031d08f9a70cc5022974179b5a3b8334b39344235f472804837ad04d483aed34
Full Changelog: v1.3.0-beta...v1.4.0-beta
API-CLI alignment and citation updated
v1.3.0-beta — the command line gives what the service gives
What changed
api_util/digital_to_json.py --markdown [PATH]writes the record as annotated Markdown, the same renderingPOST /reformatreturns withmarkdown=true.api_util/digital_to_json.py --describe [PATH]writes the per-page assessment ofPOST /describe:reader,pages,summary,stages. Its options match the endpoint's fields:--stages,--classify-pages,--no-text. The stage URLs come from--page-classification-url/--ocr-postprocess-url, or fromPAGE_CLASSIFICATION_URL/OCR_POSTPROCESS_URL. When a stage writes to the record, that write is kept under the same guard as in the service, and the record file holds it.- The CLI and the service run the same code: the new
service/describe.pyholds the stage loop, the guard, the Markdown view and the reader info, and imports no FastAPI. api_util/doc_to_visual_md.pynow takes all eight born-digital kinds, recognised by content as in/reformat. Before, it took only.pdfand.docx, by extension.
Kept apart, and why
- The Docling engine is CLI-only: a local build with model weights.
- The run's paradata is a file on the command line (
--paradata-dir) and aCreateActionin the service's answer.
Compatibility
- The API does not change; the OpenAPI spec changes only its version.
tests/test_cli_service_parity.pychecks that the CLI and the service give the same record, Markdown and assessment for the same file.
Citation
CITATION.cffcites the record schema as version 1.0 (2026-09-25), with the hub's authors.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.3.0-beta=sha256:93ee77771bf504bc4cc62895e33ae461c56fd1761f0fd66b3a0304ae71029ffeghcr.io/ufal/atrium-digital-convert-api:1.3.0-beta=sha256:2902ba30f707e86ef01bfe7d4e5a5885fbe72c2a3bded2b7db788322e8251c01
Full Changelog: v1.2.2-beta...v1.3.0-beta
Minor edits for setup config and algo logic
The record flags pages that lose letters (#5)
- A
digitalpage whose body words lose a letter to a glyph without a Unicode mapping, in 5 % of its words or more (ROUTE_DAMAGED_WORD_SHARE), is nowneeds_ocrin the record, with the damaged words counted in the reason, and its text layer isgarbled./describesends it to OCR whatever its page type. - Uploads up to 250 MB by default. A deployment can raise
MAX_UPLOAD_MB; above 250 MB memory, concurrency and timeouts are not yet tested.
The release certifies its images (ufal/atrium-project#81)
- A tag builds the images first: nothing is pushed or promoted unless the
-apiimage passes its probe. The release attachesimages.json(with its.sha256), naming each image by digest.
Documentation (ufal/atrium-project#79)
- The README follows the ATRIUM layout; the details moved to
docs/.
Images this release certifies (images.json, atrium-project#81):
ghcr.io/ufal/atrium-digital-convert-digital:1.2.2-beta=sha256:5ac5d98a9cd97ba318be2b6509dc2142e80d5f2e5840ee4aed174b0318eac5faghcr.io/ufal/atrium-digital-convert-api:1.2.2-beta=sha256:7f48377dd7df2162cf4f2704da26cf96bc6fd18036265534943faec7b6e76758
Full Changelog: v1.2.1-beta...v1.2.2-beta
Minor service changes
What's new
service/README.mdis an operator runbook (ufal/atrium-project#77): start from the published image, models and caches (none in the image; LibreOffice for DOC/XLS; the two stage services/describecalls), what to do for every error, partial results (needs_ocr,reacquire_pages, stage statuses), known limits.
No code change since v1.2.0-beta; the spec changes only its version.
Full Changelog: v1.2.0-beta...v1.2.1-beta
Minor edits for sections titles and quality stage
digital-convert v1.2.0-beta — bold section titles, and an honest quality stage
- Headings: a bold line at the body size that reads as a title is a heading, one level below the size-based ones (2–100 characters, no trailing comma, semicolon or hyphen, not a list item, at most two bold lines in a row, not repeated on three pages). New setting
BOLD_HEADING_MAX_SHARE(0.25): documents with more bold body-size text than that are left alone; 0 switches the rule off. /describe: lines ocr-postprocess could not score keep the converter's values, their page's score is recomputed from the scored lines, and the stage readspartialwith the newlines_unscored(before: stored asTrash, stageok).- Report: a page's
quality.score,langandlines_scoredcount only the lines the model judged, not itsNon-text/Emptyplaceholders. - Missing from v1.1.2-beta's notes: a page whose only text is its running header or footer over a drawing is classified by default and routed by its page type, so excavation plans with labels in the image go to OCR.
Additive: the spec stays compatible.
Full Changelog: v1.1.2-beta...v1.2.0-beta