Pages so large that the computed DPI hit the floor of 1 were still supersampled to 2 DPI, quadrupling the pixel count pdftoppm had to render for an already oversized page. Supersampling is now skipped at that floor.
The downsample test also did not prove the factor was applied, because the final 500x5000 clamp hid it in most cases. A 900x1200 render now has to come out at 450x600, which only happens when the downsample runs. The comments on the DPI cap and the clamp are reworded to describe the supersampled pipeline accurately.
Rendering page 1 straight at the thumbnail density left text slightly soft, since pdftoppm antialiases at the final size.
The page is now rendered at twice the computed DPI and downsampled by that factor with Lanczos before the existing 500x5000 clamp, which gives crisper text at about the same file size for a negligible memory cost. When the page geometry cannot be read, the fixed 150 DPI fallback is kept as a single unsupersampled render, because doubling a render that is not bounded by the thumbnail size would quadruple its pixel count. The qpdf repair path shares the same render step and gets the same behavior.
Pillow raises DecompressionBombError, which is not an OSError, so an
enormous render from the unknown-geometry fallback DPI skipped the
default thumbnail fallback. It is now converted to a ParseError like other
encoding failures, with a test covering both error types.
CI did not install qpdf, which the new thumbnail repair tests need, so it
is added to the backend package list. The setup docs now list poppler-utils
as used for thumbnail generation and no longer claim PDF thumbnails fall
back to Ghostscript when the ImageMagick PDF policy is not enabled.
The existing fallback test fails the first rasterization artificially and only proves the retry plumbing. Nothing showed that a PDF which pdftoppm genuinely cannot read is repaired by qpdf into a real rendered thumbnail, and nothing covered the case where both the render and the repair fail.
Add a test which builds a PDF with its cross reference table and trailer cut off, confirms pdftoppm rejects it, and checks that the thumbnail produced via the qpdf copy is a real 500px page while the original file is untouched. Add a parametrized test for the double failure, with qpdf itself failing and with the repaired copy still unrenderable, asserting the result is a copy of the default thumbnail.
The new pdftoppm thumbnail path capped its computed DPI at 72, on the
assumption that the old "-scale 500x5000>" never enlarged a page past its
natural size. The old pipeline actually rendered at 300 DPI before
shrinking, so any page wider than about 120pt used to produce a 500px wide
thumbnail, while the 72 DPI cap made A5 pages, receipts and other small
documents come out noticeably narrower. Rounding the DPI to the nearest
integer could also land a few pixels short of 500px (a landscape Letter page
rendered 495px wide at 45 DPI), and the Pillow clamp only ever shrinks.
The DPI is now capped at 300 to match the old render density, and rounded
up so the render always lands at or just above the target, letting the
shrink-only clamp trim it to exactly 500px. The page size helper also no
longer logs a full traceback for every encrypted or corrupt PDF, matching
the page count helper.
PDF thumbnails were produced by handing the original PDF to ImageMagick's
convert, which delegates PDF handling to Ghostscript. When that failed, a
fallback invoked gs directly and then ran convert a second time just to
encode the WebP, so a single thumbnail could take three subprocess calls
through two general purpose tools for what is only "render page one small".
The first page is now rasterized with Poppler's pdftoppm, which is already
installed for pdftotext, at a DPI computed from the page's own CropBox and
effective rotation as read by pikepdf. The DPI is chosen so the page fits
500x5000 pixels in a single render and is capped at 72 so small pages are
never enlarged, matching the previous shrink-only scale. Pillow then
flattens any alpha onto white, applies a no-enlarge safety clamp and saves
the WebP in process. If the geometry cannot be read, a fixed 150 DPI is
used and the clamp keeps the output in bounds.
The Ghostscript fallback is replaced with a qpdf repair and retry: the PDF
is copied, repaired in place with qpdf (treating its "repaired with
warnings" exit status as success), and rasterized again. If that also
fails, the default thumbnail is used as before.
* refactor: switch consumer and callers to ParserRegistry (Phase 4)
Replace all Django signal-based parser discovery with direct registry
calls. Removes `_parser_cleanup`, `parser_is_new_style` shims, and all
old-style isinstance checks. All parser instantiation now uses the
`with parser_class() as parser:` context manager pattern.
- documents/parsers.py: delegate to get_parser_registry(); drop lru_cache
- documents/consumer.py: use registry + context manager; remove shims
- documents/tasks.py: same pattern
- documents/management/commands/document_thumbnails.py: same pattern
- documents/views.py: get_metadata uses context manager
- documents/checks.py: use get_parser_registry().all_parsers()
- paperless/parsers/registry.py: add all_parsers() public method
- tests: update mocks to target documents.consumer.get_parser_class_for_mime_type
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* refactor: drop get_parser_class_for_mime_type; callers use registry directly
All callers now call get_parser_registry().get_parser_for_file() with
the actual filename and path, enabling score() to use file extension
hints. The MIME-only helper is removed.
- consumer.py: passes self.filename + self.working_copy
- tasks.py: passes document.original_filename + document.source_path
- document_thumbnails.py: same pattern
- views.py: passes Path(file).name + Path(file)
- parsers.py: internal helpers inline the registry call with filename=""
- test_parsers.py: drop TestParserDiscovery (was testing mock behavior);
TestParserAvailability uses registry directly
- test_consumer.py: mocks switch to documents.consumer.get_parser_registry
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* refactor: remove document_consumer_declaration signal infrastructure
Remove the document_consumer_declaration signal that was previously used
for parser registration. Each parser app no longer connects to this signal,
and the signal declaration itself has been removed from documents/signals.
Changes:
- Remove document_consumer_declaration from documents/signals/__init__.py
- Remove ready() methods and signal imports from all parser app configs
- Delete signal shim files (signals.py) from all parser apps:
- paperless_tesseract/signals.py
- paperless_text/signals.py
- paperless_tika/signals.py
- paperless_mail/signals.py
- paperless_remote/signals.py
Parser discovery now happens exclusively through the ParserRegistry
system introduced in the previous refactor phases.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* refactor: remove empty paperless_text and paperless_tika Django apps
After parser classes were moved to paperless/parsers/ in the plugin
refactor, these Django apps contained only empty AppConfig classes
with no models, views, tasks, migrations, or other functionality.
- Remove paperless_text and paperless_tika from INSTALLED_APPS
- Delete empty app directories entirely
- Update pyproject.toml test exclusions
- Clean stale mypy baseline entries for moved parser files
paperless_remote app is retained as it contains meaningful system
checks for Azure AI configuration.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Moves the checks and tests to the main application and removes the old applications
* Adds a comment to satisy Sonar
* refactor: remove automatic log_summary() call from get_parser_registry()
The summary was logged once per process, causing it to appear repeatedly
during Docker startup (management commands, web server, each Celery
worker subprocess). External parsers are already announced individually
at INFO when discovered; the full summary is redundant noise.
log_summary() is retained on ParserRegistry for manual/debug use.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Cleans up the duplicate test file/fixture
* Fixes a race condition where webserver threads could race to populate the registry
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* Move tesseract parser, tests, and samples to paperless.parsers
Relocates files in preparation for the Phase 3 Protocol-based parser
refactor, preserving full git history via rename.
- src/paperless_tesseract/parsers.py -> src/paperless/parsers/tesseract.py
- src/paperless_tesseract/tests/test_parser.py -> src/paperless/tests/parsers/test_tesseract_parser.py
- src/paperless_tesseract/tests/test_parser_custom_settings.py -> src/paperless/tests/parsers/test_tesseract_custom_settings.py
- src/paperless_tesseract/tests/samples/* -> src/paperless/tests/samples/tesseract/
- Moves RUF001 suppression from broad per-file pyproject.toml ignore to inline noqa comments on the two affected lines
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Refactor RasterisedDocumentParser to ParserProtocol interface
- Add RasterisedDocumentParser to registry.register_defaults()
- Update parser class: remove DocumentParser inheritance, add Protocol
class attrs/classmethods/properties, context-manager lifecycle
- Add read_file_handle_unicode_errors() to shared parsers/utils.py
- Replace inline unicode-error-handling with shared utility call
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Update tesseract signals.py to import from new parser location
RasterisedDocumentParser moved to paperless.parsers.tesseract; update
the lazy import in signals.get_parser so the signal-based consumer
declaration continues to work during the registry transition. Pop
logging_group and progress_callback kwargs for constructor compatibility.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* tests: rewrite test_tesseract_parser to pytest style with typed fixtures
- Converts all tests from Django TestCase to pytest-style classes
- Adds tesseract_samples_dir, null_app_config, tesseract_parser, and
make_tesseract_parser fixtures in conftest.py; all DB-free except
TestOcrmypdfParameters which uses @pytest.mark.django_db
- Defines MakeTesseractParser type alias in conftest.py for autocomplete
- Fixes FBT001 (boolean positional args) by making bool params
keyword-only with * separator in parametrize test signatures
- Adds type annotations to all fixture parameters for IDE support
- Uses pytest.param(..., id="...") throughout; pytest-mock for patching
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(types): fully annotate paperless/parsers/tesseract.py
Fixes all mypy and pyrefly errors in the new parser file:
- Add missing type annotations to is_image, has_alpha, get_dpi,
calculate_a4_dpi, construct_ocrmypdf_parameters, post_process_text
- Narrow Path-only (no str) for image helper args; convert to str when
building list[str] args for run_subprocess
- Annotate ocrmypdf_args as dict[str, Any] so operator expressions on
its values type-check and ocrmypdf.ocr(**args) resolves cleanly
- Declare text: str | None = None at top of extract_text to unify
all assignments to the same type across both branches
- Import Any from typing
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fixes isort
* fix: add RasterisedDocumentParser to new-style parser shim checks
The new RasterisedDocumentParser uses __enter__/__exit__ for resource
management instead of cleanup(). Update all existing new-style shims to
include it in the isinstance checks:
- documents/consumer.py: _parser_cleanup(), parser_is_new_style
- documents/tasks.py: parser_is_new_style, finally cleanup branch
(also adds RemoteDocumentParser which was missing from the latter)
- documents/management/commands/document_thumbnails.py: adds new-style
handling from scratch (enter/exit + 2-arg get_thumbnail signature)
Fix stale import paths in three test files that were still importing
from paperless_tesseract.parsers instead of paperless.parsers.tesseract.
Fix two registry tests that used application/pdf as a proxy for "no
handler" — now that RasterisedDocumentParser is registered, PDF always
has a handler, so switch to a truly unsupported MIME type.
Signal infrastructure and shims remain intact; this is plumbing only.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* One missed import (cherry pick?)
* Adds a no cover for a special case of handling unicode errors in PDF metadata
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* Chore: move Tika parser and tests to paperless/
Move TikaDocumentParser and its tests to the canonical parser package
location, matching the pattern established for TextDocumentParser:
- src/paperless_tika/parsers.py → src/paperless/parsers/tika.py
- src/paperless_tika/tests/test_tika_parser.py → src/paperless/tests/parsers/test_tika_parser.py
- src/paperless_tika/tests/samples/ → src/paperless/tests/samples/tika/
Merge tika fixtures (tika_parser, sample_odt_file, sample_docx_file,
sample_doc_file, sample_broken_odt) into the shared parsers conftest.
Remove the now-empty src/paperless_tika/tests/conftest.py.
Content is unchanged — this commit is rename-only so git history is
preserved on the moved files.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Feature: Phase 3 — migrate TikaDocumentParser to ParserProtocol
Refactor TikaDocumentParser to satisfy ParserProtocol without subclassing
the legacy DocumentParser ABC:
- Add ClassVars: name, version, author, url
- Add supported_mime_types() classmethod (12 Office/ODF/RTF MIME types)
- Add score() classmethod — returns None when TIKA_ENABLED is False, 10 otherwise
- can_produce_archive = False (PDF is for display, not an OCR archive)
- requires_pdf_rendition = True (Office formats need PDF for browser display)
- __enter__/__exit__ via ExitStack: TikaClient opened once per parser
lifetime and shared across parse() and extract_metadata() calls
- extract_metadata() falls back to a short-lived TikaClient when called
outside a context manager (legacy view-layer metadata path)
- _convert_to_pdf() uses OutputTypeConfig() to honour the database-stored
ApplicationConfiguration before falling back to the env-var setting
- Rename convert_to_pdf → _convert_to_pdf (private helper)
Update paperless_tika/signals.py shim to import from the new module path
and drop the legacy logging_group/progress_callback kwargs.
Update documents/consumer.py to extend the existing TextDocumentParser
special cases to also cover TikaDocumentParser (parse/get_thumbnail
signatures, __exit__ cleanup).
Add TestTikaParserRegistryInterface (7 tests) covering score(), properties,
and ParserProtocol isinstance check. Update existing tests to use the new
accessor API (get_text, get_date, get_archive_path, _convert_to_pdf).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: update remaining imports and move live Tika tests after parser migration
- src/documents/tests/test_parsers.py: import TikaDocumentParser from
paperless.parsers.tika (old paperless_tika.parsers no longer exists)
- git mv paperless_tika/tests/test_live_tika.py →
paperless/tests/parsers/test_live_tika.py to co-locate all Tika tests
with the parser; update import and replace old attribute API
(tika_parser.text/.archive_path) with accessor methods
(get_text/get_archive_path)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: satisfy mypy and pyrefly for TikaDocumentParser
Use a TYPE_CHECKING-guarded assert to narrow self._tika_client from
TikaClient | None to TikaClient at the point of use in parse(). The
assert is visible to type checkers (TYPE_CHECKING=True) so both mypy
and pyrefly accept the subsequent attribute accesses without error;
at runtime TYPE_CHECKING is False so the assert never executes and no
ruff S101 suppression is required.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* Fix: require context manager for TikaDocumentParser; clean up client lifecycle
- consumer.py: call __enter__ for new-style parsers so _tika_client and
_gotenberg_client are set before parse() is invoked
- views.py: use `with parser` (via nullcontext for old-style parsers) in
get_metadata so extract_metadata always runs inside a context manager
- tika.py: GotenbergClient added to ExitStack alongside TikaClient;
inline client creation removed from extract_metadata and _convert_to_pdf;
__exit__ uses ExitStack.close() instead of __exit__ pass-through
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
stumpylog
dependabot[bot]andTrenton Holmes
Sebastian Steinbeißer
Trenton Holmes
kpj
jonaswinkler