Fix: consolidate and extend unicode NFC normalization for filenames and matching

Consolidates the scattered unicodedata.normalize("NFC", ...) calls introduced
by my earlier filename/path normalization work into a single
documents.utils.normalize_unicode() helper, and closes several gaps where
NFD-normalized filenames could still slip through unnormalized:

- Document.get_public_filename() now normalizes, fixing exported filenames
  built from an NFD title/correspondent name (the default, non-format export
  path was not covered by the earlier fix).
- DocumentViewSet.update_version() now normalizes the uploaded filename,
  matching PostDocumentView's existing behavior.
- ConsumerPlugin normalizes self.filename once at consumption time, covering
  the title fallback and Document.original_filename for every document
  source (consume folder, mail, API, barcode splits).
- Workflow trigger and mail rule filename/path matching (documents/matching.py,
  paperless_mail/mail.py) now normalize the document/attachment side before
  comparing, so an NFD filename matches an NFC-typed filter pattern instead of
  silently failing to match.
- WorkflowTriggerSerializer and MailRuleSerializer normalize filter_filename/
  filter_path and the attachment include/exclude patterns once at write time,
  so the read side isn't re-normalizing an already-canonical value on every
  match.

src/documents/bulk_edit.py

@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import logging
 import tempfile
-import uuid
 from pathlib import Path
 from typing import TYPE_CHECKING
 from typing import Literal
@@ -380,7 +379,7 @@ def delete(doc_ids: list[int]) -> Literal["OK"]:
         )
         delete_ids = list({*doc_ids, *version_ids})
 
-        Document.objects.filter(id__in=delete_ids).delete(transaction_id=uuid.uuid4())
+        Document.objects.filter(id__in=delete_ids).delete()
 
         from documents.search import get_backend
 

src/documents/consumer.py

@@ -52,6 +52,7 @@ from documents.templating.workflows import parse_w_workflow_placeholders
 from documents.utils import compute_checksum
 from documents.utils import copy_basic_file_stats
 from documents.utils import copy_file_with_basic_stats
+from documents.utils import normalize_unicode
 from documents.utils import run_subprocess
 from paperless.config import OcrConfig
 from paperless.config import RemoteOCRConfig
@@ -201,7 +202,9 @@ class ConsumerPluginMixin:
 
         self.renew_logging_group()
 
-        self.filename = self.metadata.filename or self.input_doc.original_file.name
+        self.filename = normalize_unicode(
+            self.metadata.filename or self.input_doc.original_file.name,
+        )
 
     def _send_progress(
         self,

src/documents/management/commands/document_consumer.py

@@ -156,15 +156,6 @@ class FileStabilityTracker:
                     logger.debug(f"File disappeared during stability check: {path}")
                 continue
 
-            # Stable, but empty: some scanners create a zero byte placeholder
-            # and only write the page some time later. Consuming it now can
-            # only fail  so drop it and let the writer's next event
-            # (or the periodic rescan) bring it back once it has content
-            if not tracked.last_size:
-                to_remove.append(path)
-                logger.debug("Ignoring stable but empty file: %s", path)
-                continue
-
             # File is stable, we can return it
             to_yield.append(path)
             logger.info(f"File is stable: {path}")

src/documents/matching.py

@@ -21,6 +21,7 @@ from documents.models import Workflow
 from documents.models import WorkflowTrigger
 from documents.permissions import permitted_object_ids
 from documents.regex import safe_regex_search
+from documents.utils import normalize_unicode
 
 if TYPE_CHECKING:
     from django.db.models import QuerySet
@@ -311,11 +312,12 @@ def consumable_document_matches_workflow(
         trigger_matched = False
 
     # Document filename vs trigger filename
+    document_filename = normalize_unicode(document.original_file.name)
     if (
         trigger.filter_filename is not None
         and len(trigger.filter_filename) > 0
         and not fnmatch(
-            document.original_file.name.lower(),
+            document_filename.lower(),
             trigger.filter_filename.lower(),
         )
     ):
@@ -328,10 +330,12 @@ def consumable_document_matches_workflow(
     # Document path vs trigger path
 
     # Use the original_path if set, else us the original_file
-    match_against = (
+    match_against = normalize_unicode(
-        document.original_path
+        str(
-        if document.original_path is not None
+            document.original_path
-        else document.original_file
+            if document.original_path is not None
+            else document.original_file,
+        ),
     )
 
     if (
@@ -536,7 +540,7 @@ def existing_document_matches_workflow(
         and len(trigger.filter_filename) > 0
         and document.original_filename is not None
         and not fnmatch(
-            document.original_filename.lower(),
+            normalize_unicode(document.original_filename).lower(),
             trigger.filter_filename.lower(),
         )
     ):

src/documents/models.py

@@ -1,5 +1,4 @@
 import datetime
-import uuid
 from pathlib import Path
 from typing import Final
 
@@ -28,6 +27,7 @@ from django_softdelete.models import SoftDeleteModel
 
 from documents.data_models import DocumentSource
 from documents.parsers import get_default_file_extension
+from documents.utils import normalize_unicode
 
 
 class ModelWithOwner(models.Model):
@@ -468,7 +468,7 @@ class Document(SoftDeleteModel, ModelWithOwner):  # type: ignore[django-manager-
         context_document = (
             self.root_document if self.root_document_id is not None else self
         )
-        result = str(context_document)
+        result = normalize_unicode(str(context_document))
 
         if counter:
             result += f"_{counter:02}"
@@ -515,20 +515,13 @@ class Document(SoftDeleteModel, ModelWithOwner):  # type: ignore[django-manager-
     def delete(
         self,
         *args,
-        transaction_id=None,
         **kwargs,
     ):
-        # Versions must share the root's transaction ID so they are restored
+        # If deleting a root document, move all its versions to trash as well.
-        # together by django-softdelete.
-        if transaction_id is None:
-            transaction_id = uuid.uuid4()
         if self.root_document_id is None:
-            Document.objects.filter(root_document=self).delete(
+            Document.objects.filter(root_document=self).delete()
-                transaction_id=transaction_id,
-            )
         return super().delete(
             *args,
-            transaction_id=transaction_id,
             **kwargs,
         )
 

src/documents/serialisers.py

@@ -87,6 +87,7 @@ from documents.regex import validate_regex_pattern
 from documents.templating.filepath import validate_filepath_template_and_render
 from documents.templating.utils import convert_format_str_to_template_format
 from documents.templating.workflows import validate_workflow_template
+from documents.utils import normalize_unicode
 from documents.validators import uri_validator
 from documents.validators import url_validator
 from documents.versioning import sort_versions_newest_first
@@ -3120,6 +3121,13 @@ class WorkflowTriggerSerializer(serializers.ModelSerializer[WorkflowTrigger]):
         ):
             attrs["filter_path"] = None
 
+        # Normalize once at write time, since these are matched against many
+        # documents but edited rarely
+        if attrs.get("filter_filename") is not None:
+            attrs["filter_filename"] = normalize_unicode(attrs["filter_filename"])
+        if attrs.get("filter_path") is not None:
+            attrs["filter_path"] = normalize_unicode(attrs["filter_path"])
+
         if (
             "filter_custom_field_query" in attrs
             and attrs["filter_custom_field_query"] is not None

src/documents/templating/filepath.py

@@ -1,7 +1,6 @@
 import logging
 import os
 import re
-import unicodedata
 from collections.abc import Iterable
 from pathlib import PurePath
 
@@ -26,6 +25,7 @@ from documents.templating.environment import _template_environment
 from documents.templating.filters import format_datetime
 from documents.templating.filters import get_cf_value
 from documents.templating.filters import localize_date
+from documents.utils import normalize_unicode
 
 logger = logging.getLogger("paperless.templating")
 
@@ -42,7 +42,7 @@ class FilePathTemplate(Template):
             3. Removing extra spaces before and after forward slashes
             4. Preserving spaces in other parts of the path
             """
-            value = unicodedata.normalize("NFC", value)
+            value = normalize_unicode(value)
             value = value.replace("\n", "").replace("\r", "")
             value = re.sub(r"\s*/\s*", "/", value)
 
@@ -184,17 +184,17 @@ def get_basic_metadata_context(
     """
     return {
         "title": pathvalidate.sanitize_filename(
-            unicodedata.normalize("NFC", document.title),
+            normalize_unicode(document.title),
             replacement_text="-",
         ),
         "correspondent": pathvalidate.sanitize_filename(
-            unicodedata.normalize("NFC", document.correspondent.name),
+            normalize_unicode(document.correspondent.name),
             replacement_text="-",
         )
         if document.correspondent
         else no_value_default,
         "document_type": pathvalidate.sanitize_filename(
-            unicodedata.normalize("NFC", document.document_type.name),
+            normalize_unicode(document.document_type.name),
             replacement_text="-",
         )
         if document.document_type
@@ -205,8 +205,7 @@ def get_basic_metadata_context(
         "owner_username": document.owner.username
         if document.owner
         else no_value_default,
-        "original_name": unicodedata.normalize(
+        "original_name": normalize_unicode(
-            "NFC",
             PurePath(document.original_filename).with_suffix("").name,
         )
         if document.original_filename
@@ -275,12 +274,12 @@ def get_tags_context(tags: Iterable[Tag]) -> dict[str, str | list[str]]:
     return {
         "tag_list": pathvalidate.sanitize_filename(
             ",".join(
-                sorted(unicodedata.normalize("NFC", tag.name) for tag in tags),
+                sorted(normalize_unicode(tag.name) for tag in tags),
             ),
             replacement_text="-",
         ),
         # Assumed to be ordered, but a template could loop through to find what they want
-        "tag_name_list": [unicodedata.normalize("NFC", x.name) for x in tags],
+        "tag_name_list": [normalize_unicode(x.name) for x in tags],
     }
 
 
@@ -307,7 +306,7 @@ def get_custom_fields_context(
             CustomField.FieldDataType.LONG_TEXT,
         }:
             value = pathvalidate.sanitize_filename(
-                unicodedata.normalize("NFC", field_instance.value),
+                normalize_unicode(field_instance.value),
                 replacement_text="-",
             )
         elif (
@@ -316,8 +315,7 @@ def get_custom_fields_context(
         ):
             options = field_instance.field.extra_data["select_options"]
             value = pathvalidate.sanitize_filename(
-                unicodedata.normalize(
+                normalize_unicode(
-                    "NFC",
                     next(
                         option["label"]
                         for option in options
@@ -330,7 +328,7 @@ def get_custom_fields_context(
             value = field_instance.value
         field_data["custom_fields"][
             pathvalidate.sanitize_filename(
-                unicodedata.normalize("NFC", field_instance.field.name),
+                normalize_unicode(field_instance.field.name),
                 replacement_text="-",
             )
         ] = {

src/documents/tests/test_api_trash.py

@@ -207,65 +207,3 @@ class TestTrashAPI(DirectoriesMixin, APITestCase):
         )
         self.assertEqual(resp.status_code, status.HTTP_400_BAD_REQUEST)
         self.assertIn("have not yet been deleted", resp.data["documents"][0])
-
-    def _make_versioned_document(self) -> tuple[Document, list[Document]]:
-        root = Document.objects.create(
-            title="root",
-            content="root-content",
-            checksum="root",
-            mime_type="application/pdf",
-        )
-        versions = [
-            Document.objects.create(
-                title=f"v{index}",
-                content=f"v{index}-content",
-                checksum=f"v{index}",
-                mime_type="application/pdf",
-                root_document=root,
-                version_index=index,
-            )
-            for index in range(1, 3)
-        ]
-        return root, versions
-
-    def test_api_trash_restore_document_restores_its_versions(self) -> None:
-        """
-        GIVEN:
-            - Existing document with two versions
-        WHEN:
-            - API request to delete the document
-            - API request to restore it from the trash
-        THEN:
-            - Only the document itself is listed in the trash
-            - A version cannot be restored without its root
-            - The document is restored together with all of its versions
-        """
-        root, versions = self._make_versioned_document()
-
-        self.client.force_login(user=self.user)
-        self.client.delete(f"/api/documents/{root.pk}/")
-        self.assertEqual(Document.deleted_objects.count(), 3)
-
-        resp = self.client.get("/api/trash/")
-        self.assertEqual(resp.status_code, status.HTTP_200_OK)
-        self.assertEqual(resp.data["count"], 1)
-        self.assertEqual(resp.data["results"][0]["id"], root.pk)
-
-        # A version cannot be restored while its root remains in the trash.
-        resp = self.client.post(
-            "/api/trash/",
-            {"action": "restore", "documents": [versions[0].pk]},
-        )
-        self.assertEqual(resp.status_code, status.HTTP_400_BAD_REQUEST)
-        self.assertIn("Restore the root document", resp.data["documents"][0])
-
-        resp = self.client.post(
-            "/api/trash/",
-            {"action": "restore", "documents": [root.pk]},
-        )
-        self.assertEqual(resp.status_code, status.HTTP_200_OK)
-        self.assertEqual(Document.deleted_objects.count(), 0)
-        self.assertCountEqual(
-            Document.objects.filter(root_document=root).values_list("id", flat=True),
-            [version.pk for version in versions],
-        )

src/documents/tests/test_api_update_version_nfc.py

@@ -0,0 +1,69 @@
+import unicodedata
+from typing import TYPE_CHECKING
+from unittest import mock
+
+import celery.result
+import pytest
+from django.core.files.uploadedfile import SimpleUploadedFile
+
+from documents.models import Document
+
+if TYPE_CHECKING:
+    from documents.data_models import ConsumableDocument
+
+
+@pytest.fixture()
+def consume_file_mock():
+    with mock.patch("documents.tasks.consume_file.apply_async") as m:
+        m.return_value = celery.result.AsyncResult(id="test-task-id")
+        yield m
+
+
+@pytest.fixture()
+def directories(tmp_path, settings, _media_settings):
+    scratch = tmp_path / "scratch"
+    scratch.mkdir()
+    settings.SCRATCH_DIR = scratch
+    return scratch
+
+
+@pytest.mark.django_db
+class TestUpdateVersionNFCNormalization:
+    def test_nfd_filename_normalized_to_nfc(
+        self,
+        admin_client,
+        consume_file_mock: mock.MagicMock,
+        directories,
+    ):
+        """Uploaded new-version file with NFD filename must have its temp name stored as NFC."""
+        document = Document.objects.create(
+            title="Test",
+            content="content",
+            checksum="checksum",
+            mime_type="application/pdf",
+        )
+
+        nfd = unicodedata.normalize("NFD", "Rechnung März.pdf")
+        nfc = unicodedata.normalize("NFC", "Rechnung März.pdf")
+
+        assert nfd != nfc
+
+        uploaded = SimpleUploadedFile(
+            nfd,
+            b"%PDF-1.4 test",
+            content_type="application/pdf",
+        )
+        response = admin_client.post(
+            f"/api/documents/{document.pk}/update_version/",
+            {"document": uploaded},
+        )
+
+        assert response.status_code == 200
+
+        task_kwargs = consume_file_mock.call_args.kwargs["kwargs"]
+        input_doc: ConsumableDocument = task_kwargs["input_doc"]
+
+        assert input_doc.original_file.name == nfc, (
+            f"Expected NFC filename {nfc!r}, got {input_doc.original_file.name!r}"
+        )
+        assert unicodedata.is_normalized("NFC", input_doc.original_file.name)

src/documents/tests/test_bulk_edit.py

@@ -392,11 +392,6 @@ class TestBulkEdit(DirectoriesMixin, TestCase):
         self.assertFalse(Document.objects.filter(id=self.doc1.id).exists())
         self.assertFalse(Document.objects.filter(id=version.id).exists())
 
-        Document.deleted_objects.get(id=self.doc1.id).restore(strict=False)
-
-        self.assertTrue(Document.objects.filter(id=self.doc1.id).exists())
-        self.assertTrue(Document.objects.filter(id=version.id).exists())
-
     def test_delete_version_document_keeps_root(self) -> None:
         version = Document.objects.create(
             checksum="A-v1",

src/documents/tests/test_document_model.py

@@ -110,7 +110,7 @@ class TestDocument(TestCase):
             checksum="checksum",
             mime_type="application/pdf",
         )
-        version = Document.objects.create(
+        Document.objects.create(
             root_document=root,
             correspondent=root.correspondent,
             title="Version",
@@ -124,10 +124,6 @@ class TestDocument(TestCase):
         self.assertEqual(Document.objects.count(), 0)
         self.assertEqual(Document.deleted_objects.count(), 2)
 
-        root.restore(strict=False)
-
-        self.assertTrue(Document.objects.filter(pk=version.pk).exists())
-
     def test_file_name(self) -> None:
         doc = Document(
             mime_type="application/pdf",

src/documents/tests/test_get_public_filename_nfc.py

@@ -0,0 +1,48 @@
+import unicodedata
+from datetime import date
+
+import pytest
+
+from documents.models import Correspondent
+from documents.models import Document
+
+
+@pytest.mark.django_db
+class TestGetPublicFilenameNfc:
+    def test_normalizes_nfd_title_to_nfc(self) -> None:
+        nfd_title = unicodedata.normalize("NFD", "Gehaltserhöhung")
+        assert not unicodedata.is_normalized("NFC", nfd_title)
+
+        doc = Document(
+            mime_type="application/pdf",
+            title=nfd_title,
+            created=date(2025, 10, 17),
+        )
+
+        result = doc.get_public_filename()
+
+        assert unicodedata.is_normalized("NFC", result)
+        assert (
+            result
+            == "2025-10-17 "
+            + unicodedata.normalize(
+                "NFC",
+                nfd_title,
+            )
+            + ".pdf"
+        )
+
+    def test_normalizes_nfd_correspondent_name_to_nfc(self) -> None:
+        nfd_name = unicodedata.normalize("NFD", "Müller GmbH")
+        correspondent = Correspondent.objects.create(name=nfd_name)
+
+        doc = Document.objects.create(
+            mime_type="application/pdf",
+            title="Rechnung",
+            created=date(2025, 10, 17),
+            correspondent=correspondent,
+        )
+
+        result = doc.get_public_filename()
+
+        assert unicodedata.is_normalized("NFC", result)

src/documents/tests/test_management_consumer.py

@@ -136,23 +136,6 @@ def wait_for_mock_call(
     return False
 
 
-def sleep_past_stability(
-    owner: FileStabilityTracker | ConsumerThread,
-    *,
-    windows: float = 1.5,
-) -> None:
-    """
-    Block until a tracked file's stability window has certainly elapsed.
-
-    Args:
-        owner: The tracker, or the consumer thread running one, whose
-               configured stability delay sets the wait.
-        windows: How many stability windows to wait, giving slop for a slow
-                 or loaded test runner.
-    """
-    sleep(owner.stability_delay * windows)
-
-
 class TestTrackedFile:
     """Tests for the TrackedFile dataclass."""
 
@@ -278,56 +261,6 @@ class TestFileStabilityTracker:
         assert len(stable) == 0
         assert stability_tracker.pending_count == 1
 
-    def test_get_stable_files_skips_empty_file(
-        self,
-        stability_tracker: FileStabilityTracker,
-        tmp_path: Path,
-    ) -> None:
-        """
-        GIVEN:
-            - A zero byte file, tracked and past its stability delay
-        WHEN:
-            - Stable files are collected
-        THEN:
-            - The file is not yielded for consumption
-            - The file is dropped from tracking rather than held, so an
-              abandoned placeholder does not keep the watch loop awake
-        """
-        empty = tmp_path / "scan.pdf"
-        empty.write_bytes(b"")
-        stability_tracker.track(empty, Change.added)
-        sleep_past_stability(stability_tracker)
-
-        stable = list(stability_tracker.get_stable_files())
-
-        assert stable == []
-        assert stability_tracker.pending_count == 0
-
-    def test_empty_file_is_yielded_once_content_arrives(
-        self,
-        stability_tracker: FileStabilityTracker,
-        tmp_path: Path,
-    ) -> None:
-        """
-        GIVEN:
-            - A zero byte file which was dropped from tracking while empty
-        WHEN:
-            - The writer fills the file and a new event re-tracks it
-        THEN:
-            - The file is yielded for consumption once it is stable
-        """
-        target = tmp_path / "scan.pdf"
-        target.write_bytes(b"")
-        stability_tracker.track(target, Change.added)
-        sleep_past_stability(stability_tracker)
-        assert list(stability_tracker.get_stable_files()) == []
-
-        target.write_bytes(b"%PDF-1.4 content")
-        stability_tracker.track(target, Change.modified)
-        sleep_past_stability(stability_tracker)
-
-        assert list(stability_tracker.get_stable_files()) == [target]
-
     def test_get_stable_files_deleted_during_check(self, temp_file: Path) -> None:
         """Test deleted file is not returned during stability check."""
         tracker = FileStabilityTracker(stability_delay=0.1)
@@ -946,51 +879,6 @@ class TestCommandWatch:
 
         mock_consume_file_delay.apply_async.assert_called()
 
-    def test_scanner_placeholder_is_not_consumed_while_empty(
-        self,
-        consumption_dir: Path,
-        sample_pdf: Path,
-        mock_consume_file_delay: MagicMock,
-        start_consumer: Callable[..., ConsumerThread],
-    ) -> None:
-        """
-        GIVEN:
-            - A scanner which creates a zero byte placeholder and only writes
-              the page some time later (GH discussion #13969)
-        WHEN:
-            - The placeholder sits untouched well past the stability delay
-            - The scanner then writes the real content
-        THEN:
-            - The empty placeholder is never queued, as it could only fail
-              with "Unsupported mime type inode/x-empty"
-            - The file is queued exactly once, when the content lands
-        """
-        thread = start_consumer(stability_delay=0.2)
-
-        target = consumption_dir / "scan.pdf"
-        target.write_bytes(b"")  # the scanner's placeholder
-
-        # Well past the stability delay: the old behaviour queued it here.
-        sleep_past_stability(thread, windows=5)
-        if thread.exception:
-            raise thread.exception
-        assert mock_consume_file_delay.apply_async.call_count == 0
-
-        shutil.copy(sample_pdf, target)  # the scanner finishes the page
-
-        assert wait_for_mock_call(
-            mock_consume_file_delay.apply_async,
-            timeout_s=5.0,
-        )
-        if thread.exception:
-            raise thread.exception
-
-        assert mock_consume_file_delay.apply_async.call_count == 1
-        queued_doc = mock_consume_file_delay.apply_async.call_args.kwargs["kwargs"][
-            "input_doc"
-        ]
-        assert queued_doc.original_file.name == "scan.pdf"
-
     def test_ignores_macos_files(
         self,
         consumption_dir: Path,

src/documents/tests/test_matching_nfc.py

@@ -0,0 +1,80 @@
+import unicodedata
+
+import pytest
+
+from documents.data_models import ConsumableDocument
+from documents.data_models import DocumentSource
+from documents.matching import consumable_document_matches_workflow
+from documents.matching import existing_document_matches_workflow
+from documents.models import Document
+from documents.models import Workflow
+from documents.models import WorkflowTrigger
+
+
+@pytest.mark.django_db
+class TestMatchingNfcNormalization:
+    def test_consumable_document_filename_nfd_matches_nfc_pattern(
+        self,
+        tmp_path,
+    ) -> None:
+        """
+        GIVEN:
+            - A file on disk whose name is NFD-normalized
+            - A workflow trigger filename filter typed as NFC
+        WHEN:
+            - The consumable document is checked against the trigger
+        THEN:
+            - It matches, because both sides are normalized before comparing
+        """
+        nfd_name = unicodedata.normalize("NFD", "Gehaltserhöhung.pdf")
+        nfc_pattern = unicodedata.normalize("NFC", "*Gehaltserhöhung*")
+        assert nfd_name != unicodedata.normalize("NFC", nfd_name)
+
+        file_path = tmp_path / nfd_name
+        file_path.write_bytes(b"%PDF-1.4 test")
+
+        document = ConsumableDocument(
+            source=DocumentSource.ConsumeFolder,
+            original_file=file_path,
+        )
+        trigger = WorkflowTrigger(
+            type=WorkflowTrigger.WorkflowTriggerType.CONSUMPTION,
+            filter_filename=nfc_pattern,
+            sources=[],
+        )
+
+        matched, reason = consumable_document_matches_workflow(document, trigger)
+
+        assert matched, reason
+
+    def test_existing_document_filename_nfd_matches_nfc_pattern(self) -> None:
+        """
+        GIVEN:
+            - A Document whose original_filename is NFD-normalized (e.g. from
+              before normalization was applied at consumption time)
+            - A workflow trigger filename filter typed as NFC
+        WHEN:
+            - The document is checked against the trigger
+        THEN:
+            - It matches, because both sides are normalized before comparing
+        """
+        nfd_name = unicodedata.normalize("NFD", "Gehaltserhöhung.pdf")
+        nfc_pattern = unicodedata.normalize("NFC", "*Gehaltserhöhung*")
+
+        document = Document.objects.create(
+            title="Test",
+            content="content",
+            checksum="checksum",
+            mime_type="application/pdf",
+            original_filename=nfd_name,
+        )
+        workflow = Workflow.objects.create(name="Test workflow", order=0)
+        trigger = WorkflowTrigger.objects.create(
+            type=WorkflowTrigger.WorkflowTriggerType.DOCUMENT_ADDED,
+            filter_filename=nfc_pattern,
+        )
+        workflow.triggers.add(trigger)
+
+        matched, reason = existing_document_matches_workflow(document, trigger)
+
+        assert matched, reason

src/documents/tests/test_normalize_unicode.py

@@ -0,0 +1,6 @@
+from documents.utils import normalize_unicode
+
+
+class TestNormalizeUnicode:
+    def test_none_passes_through(self) -> None:
+        assert normalize_unicode(None) is None

src/documents/tests/test_views.py

@@ -32,7 +32,6 @@ from documents.signals.handlers import update_llm_suggestions_cache
 from documents.tests.utils import DirectoriesMixin
 from documents.tests.utils import read_streaming_response
 from paperless.models import ApplicationConfiguration
-from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 
 
@@ -738,38 +737,6 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             get_llm_suggestion_cache(self.document.pk, backend="openai-like"),
         )
 
-    @patch("documents.views.get_ai_document_classification")
-    @override_settings(
-        AI_ENABLED=True,
-        LLM_BACKEND="openai-like",
-    )
-    def test_ai_suggestions_with_llm_provider_error(
-        self,
-        mock_get_ai_classification,
-    ) -> None:
-        mock_get_ai_classification.side_effect = LLMProviderError(
-            "confidential provider response",
-        )
-
-        self.client.force_login(user=self.user)
-        response = self.client.get(
-            f"/api/documents/{self.document.pk}/ai_suggestions/",
-        )
-
-        self.assertEqual(response.status_code, status.HTTP_502_BAD_GATEWAY)
-        self.assertEqual(
-            response.json(),
-            {
-                "ai": [
-                    "AI backend rejected the request. Check logs for details.",
-                ],
-            },
-        )
-        self.assertNotIn("confidential provider response", response.content.decode())
-        self.assertIsNone(
-            get_llm_suggestion_cache(self.document.pk, backend="openai-like"),
-        )
-
     @patch("documents.views.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,

src/documents/utils.py

@@ -1,6 +1,7 @@
 import hashlib
 import logging
 import shutil
+import unicodedata
 from collections.abc import Callable
 from collections.abc import Iterable
 from collections.abc import Iterator
@@ -31,6 +32,25 @@ def identity(iterable: Iterable[_T]) -> Iterable[_T]:
     return iterable
 
 
+def normalize_unicode(value: str | None) -> str | None:
+    """
+    Normalize a string to Unicode NFC form, or return None unchanged.
+
+    This is the single normalization pass for any user- or filesystem-supplied
+    text that ends up in a filename, path, or is compared/matched against one
+    (titles, correspondent/tag/type names, uploaded filenames, workflow and
+    mail rule filename/path filters). Composed (NFC) and decomposed (NFD)
+    forms of the same visible text are different byte sequences, which breaks
+    exact comparisons and filesystem lookups even though the text looks
+    identical. Always normalize through this function rather than calling
+    unicodedata.normalize() directly, so every call site agrees on the same
+    form.
+    """
+    if value is None:
+        return None
+    return unicodedata.normalize("NFC", value)
+
+
 class QuerySetStream(Generic[_M]):
     """Stream a QuerySet via .iterator(chunk_size=...) instead of
     materializing it (plus any prefetch caches) all at once, while still

src/documents/views.py

@@ -231,6 +231,7 @@ from documents.tasks import sanity_check
 from documents.tasks import train_classifier
 from documents.tasks import update_document_parent_tags
 from documents.utils import get_boolean
+from documents.utils import normalize_unicode
 from documents.versioning import VersionResolutionError
 from documents.versioning import annotate_effective_content
 from documents.versioning import get_latest_version_for_root
@@ -252,7 +253,6 @@ from paperless.views import StandardPagination
 from paperless_ai.ai_classifier import get_ai_document_classification
 from paperless_ai.ai_classifier import get_llm_output_language
 from paperless_ai.chat import stream_chat_with_documents
-from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 from paperless_ai.matching import extract_unmatched_names
 from paperless_ai.matching import match_correspondents_by_name
@@ -1604,22 +1604,6 @@ class DocumentViewSet(
                     {"ai": [_("AI backend request timed out.")]},
                     status=status.HTTP_503_SERVICE_UNAVAILABLE,
                 )
-            except LLMProviderError:
-                logger.exception(
-                    "AI backend rejected the request for document %s",
-                    doc.pk,
-                )
-                return Response(
-                    {
-                        "ai": [
-                            _(
-                                "AI backend rejected the request. "
-                                "Check logs for details.",
-                            ),
-                        ],
-                    },
-                    status=status.HTTP_502_BAD_GATEWAY,
-                )
             set_llm_suggestions_cache(
                 doc.pk,
                 llm_suggestions,
@@ -2085,6 +2069,7 @@ class DocumentViewSet(
 
         try:
             doc_name, doc_data = serializer.validated_data.get("document")
+            doc_name = normalize_unicode(doc_name)
             version_label = serializer.validated_data.get("version_label")
 
             t = int(mktime(datetime.now().timetuple()))
@@ -3351,7 +3336,7 @@ class PostDocumentView(GenericAPIView[Any]):
         serializer.is_valid(raise_exception=True)
 
         doc_name, doc_data = serializer.validated_data.get("document")
-        doc_name = normalize("NFC", doc_name)
+        doc_name = normalize_unicode(doc_name)
         correspondent_id = serializer.validated_data.get("correspondent")
         document_type_id = serializer.validated_data.get("document_type")
         storage_path_id = serializer.validated_data.get("storage_path")
@@ -5454,10 +5439,7 @@ class TrashView(ListModelMixin, PassUserMixin):
 
     model = Document
 
-    # A version is listed separately only when its root is not in the trash.
+    queryset = Document.deleted_objects.all()
-    queryset = Document.deleted_objects.exclude(
-        root_document_id__in=Document.deleted_objects.values("id"),
-    )
 
     def get(self, request: Request, format: str | None = None) -> Response:
         self.serializer_class = DocumentSerializer
@@ -5488,15 +5470,7 @@ class TrashView(ListModelMixin, PassUserMixin):
             return HttpResponseForbidden("Insufficient permissions")
         action = serializer.validated_data.get("action")
         if action == "restore":
-            restored = list(self.get_queryset().filter(id__in=doc_ids))
+            restored = list(Document.deleted_objects.filter(id__in=doc_ids))
-            if len(restored) != len(doc_ids):
-                raise ValidationError(
-                    {
-                        "documents": [
-                            "Restore the root document instead of one of its versions.",
-                        ],
-                    },
-                )
             for doc in restored:
                 doc.restore(strict=False)
             if restored:

src/paperless_ai/ai_classifier.py

@@ -4,24 +4,21 @@ from django.conf import settings
 from django.contrib.auth.models import User
 
 from documents.models import Document
-from documents.permissions import permitted_object_ids
+from documents.permissions import get_objects_for_user_owner_aware
-from documents.permissions import restrict_queryset_to_visible
-from documents.permissions import user_is_unrestricted
 from paperless.config import AIConfig
 from paperless_ai.base_model import ClassificationSuggestions
 from paperless_ai.base_model import TaxonomyChoiceDict
 from paperless_ai.base_model import classification_suggestions_to_model
 from paperless_ai.client import AIClient
 from paperless_ai.db import db_connection_released
+from paperless_ai.indexing import _node_document_ids
 from paperless_ai.indexing import retrieve_similar_nodes
 from paperless_ai.indexing import truncate_content
 from paperless_ai.prompts.context import ClassificationPromptContext
 from paperless_ai.prompts.context import LocalizationPromptContext
 from paperless_ai.prompts.context import RagContextPromptContext
 from paperless_ai.prompts.render import render_prompt
-from paperless_ai.taxonomy import SimilarDocument
 from paperless_ai.taxonomy import TaxonomyCandidates
-from paperless_ai.taxonomy import _node_document_weights
 from paperless_ai.taxonomy import build_taxonomy_candidates
 from paperless_ai.taxonomy import empty_taxonomy_candidates
 from paperless_ai.taxonomy import format_taxonomy_for_prompt
@@ -40,48 +37,6 @@ logger = logging.getLogger("paperless_ai.rag_classifier")
 TAXONOMY_CANDIDATE_TOP_K = 15
 
 
-def _fulltext_similar_documents(
-    document: Document,
-    user: User | None,
-    top_k: int,
-) -> list[SimilarDocument]:
-    """Rank-based fallback when no embedding backend is configured. Uses
-    Tantivy's "More Like This" (term-overlap similarity) instead of vector
-    similarity - cruder, but far better than no candidates at all.
-    more_like_this_ids returns only a ranked ID list, no scores, so weight is
-    synthesized from rank (descending from top_k) rather than claiming a
-    similarity magnitude that doesn't exist. An unrestricted user (none, or an
-    active superuser - see user_is_unrestricted) is normalized to ``None``
-    before calling, since the backend's permission filter has no superuser
-    short-circuit of its own. Results are re-checked with
-    restrict_queryset_to_visible() since Tantivy's indexed permission fields
-    lag the DB via async reindexing.
-    """
-    from documents.search import get_backend
-
-    unrestricted = user_is_unrestricted(user)
-    search_user = None if unrestricted else user
-    backend = get_backend()
-    similar_ids = backend.more_like_this_ids(
-        document.pk,
-        user=search_user,
-        limit=top_k,
-    )
-    if not unrestricted:
-        allowed_ids = set(
-            restrict_queryset_to_visible(
-                Document.objects.filter(pk__in=similar_ids),
-                user,
-                "view_document",
-            ).values_list("pk", flat=True),
-        )
-        similar_ids = [doc_id for doc_id in similar_ids if doc_id in allowed_ids]
-    return [
-        SimilarDocument(document_id=doc_id, weight=float(top_k - rank))
-        for rank, doc_id in enumerate(similar_ids)
-    ]
-
-
 def get_language_name(language_code: str) -> str:
     normalized_language_code = language_code.lower()
     for code, name in settings.LANGUAGES:
@@ -181,52 +136,43 @@ def get_taxonomy_context(
     user: User | None = None,
     max_docs: int = 5,
 ) -> tuple[TaxonomyCandidates, str]:
-    """One retrieval feeds both taxonomy candidates and RAG text context. Uses
+    """One retrieval feeds both taxonomy candidates and RAG text context.
-    vector similarity when an embedding backend is configured, otherwise
+    On any retrieval failure, degrades to empty candidates/context rather than
-    falls back to Tantivy full-text "More Like This" similarity - see
+    propagating the exception - a vector-store outage should not block
-    _fulltext_similar_documents. On any retrieval failure, degrades to empty
+    classification, only its RAG-assisted enrichment.
-    candidates/context rather than propagating the exception - neither a
-    vector-store outage nor a search-index issue should block classification,
-    only its context-assisted enrichment.
     """
-    ai_config = AIConfig()
     try:
-        if ai_config.llm_embedding_backend:
+        # None means "no restriction" to retrieve_similar_nodes. A superuser
-            # None means "no restriction" to retrieve_similar_nodes. An
+        # (like no user at all) can see every document, so skip materializing
-            # unrestricted user (no user at all, or an active superuser -- see
+        # every visible pk into a Python list and passing it through as an IN
-            # user_is_unrestricted) can see every document, so skip
+        # filter: for a large library that is a wasted quadratic scan in the
-            # materializing every visible pk into a Python list and passing it
+        # vector store at best, and past ~32,763 documents a hard
-            # through as an IN filter: for a large library that is a wasted
+        # sqlite3.OperationalError (SQLite's bound-parameter limit) at worst.
-            # quadratic scan in the vector store at best, and past ~32,763
+        # get_objects_for_user_owner_aware() would return every Document for a
-            # documents a hard sqlite3.OperationalError (SQLite's
+        # superuser anyway (guardian's own with_superuser shortcut), so this
-            # bound-parameter limit) at worst.
+        # changes nothing about which documents are considered -- only how we
-            # permitted_object_ids() has its own superuser shortcut that would
+        # get there.
-            # return every Document's id anyway, so this changes nothing about
+        visible_document_ids = (
-            # which documents are considered -- only how we get there.
+            None
-            visible_document_ids = (
+            if user is None or user.is_superuser
-                None
+            else list(
-                if user_is_unrestricted(user)
+                get_objects_for_user_owner_aware(
-                else list(permitted_object_ids(user, Document, "view_document"))
+                    user,
-            )
+                    "view_document",
-            nodes = retrieve_similar_nodes(
+                    Document,
-                document,
+                ).values_list("pk", flat=True),
-                top_k=TAXONOMY_CANDIDATE_TOP_K,
-                document_ids=visible_document_ids,
-            )
-            similar_documents = _node_document_weights(nodes)
-        else:
-            # See _fulltext_similar_documents: it applies its own permission
-            # filter via `user`, so no visible-document-id list is needed here.
-            similar_documents = _fulltext_similar_documents(
-                document,
-                user,
-                top_k=TAXONOMY_CANDIDATE_TOP_K,
             )
+        )
+        nodes = retrieve_similar_nodes(
+            document,
+            top_k=TAXONOMY_CANDIDATE_TOP_K,
+            document_ids=visible_document_ids,
+        )
 
-        candidates = build_taxonomy_candidates(similar_documents, user)
+        candidates = build_taxonomy_candidates(nodes, user)
 
-        # similar_documents is already ordered by descending weight; don't lose it.
+        # ``nodes`` are already ordered by descending vector similarity; don't lose it.
-        similar_document_ids = [s["document_id"] for s in similar_documents]
+        similar_document_ids = list(dict.fromkeys(_node_document_ids(nodes)))
         similar_documents_by_id = Document.objects.in_bulk(similar_document_ids)
         similar_docs = [
             similar_documents_by_id[document_id]
@@ -240,8 +186,8 @@ def get_taxonomy_context(
             context_blocks.append(f"TITLE: {title}\n{text}")
     except Exception:
         logger.exception(
-            "Failed to retrieve similar-document context for document %s; "
+            "Failed to retrieve RAG neighbours for document %s; continuing "
-            "continuing without taxonomy candidates or similar-document context.",
+            "without taxonomy candidates or similar-document context.",
             document.pk,
         )
         return empty_taxonomy_candidates(), ""
@@ -295,13 +241,17 @@ def get_ai_document_classification(
 ) -> ClassificationSuggestions:
     ai_config = AIConfig()
 
-    candidates, context = get_taxonomy_context(document, user)
+    if ai_config.llm_embedding_backend:
-    prompt = build_prompt_with_rag(
+        candidates, context = get_taxonomy_context(document, user)
-        document,
+        prompt = build_prompt_with_rag(
-        ai_config,
+            document,
-        candidates=candidates,
+            ai_config,
-        context=context,
+            candidates=candidates,
-    )
+            context=context,
+        )
+    else:
+        candidates = empty_taxonomy_candidates()
+        prompt = build_prompt_without_rag(document, ai_config, candidates=candidates)
 
     client = AIClient()
     # Hand the pooled DB connection back while the (slow) LLM query runs so it

src/paperless_ai/client.py

@@ -22,7 +22,6 @@ from paperless.network import validate_outbound_http_url
 from paperless_ai.base_model import ClassificationSuggestions
 from paperless_ai.base_model import DocumentClassifierSchema
 from paperless_ai.base_model import model_to_classification_suggestions
-from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 
 logger = logging.getLogger("paperless_ai.client")
@@ -133,7 +132,7 @@ class AIClient:
         from llama_index.core.llms import ChatMessage
 
         if self.settings.llm_backend == LLMBackend.OLLAMA:
-            with self._normalize_errors():
+            with self._normalize_timeouts():
                 result = self.llm.chat(
                     [ChatMessage(role="user", content=prompt)],
                     format=DocumentClassifierSchema.model_json_schema(),
@@ -154,7 +153,7 @@ class AIClient:
             content=f"{prompt}\n\n"
             f"Answer by calling the {tool.metadata.name} tool. Do not write the answer as text.",
         )
-        with self._normalize_errors():
+        with self._normalize_timeouts():
             result = self.llm.chat_with_tools(
                 tools=[tool],
                 user_msg=user_msg,
@@ -174,7 +173,7 @@ class AIClient:
         )
 
     @contextmanager
-    def _normalize_errors(self) -> Iterator[None]:
+    def _normalize_timeouts(self) -> Iterator[None]:
         try:
             yield
         except httpx.TimeoutException as exc:
@@ -182,23 +181,8 @@ class AIClient:
         except Exception as exc:
             if self._is_openai_timeout(exc):
                 raise LLMTimeoutError from exc
-            if self._is_provider_error(exc):
-                raise LLMProviderError from exc
             raise
 
-    def _is_provider_error(self, exc: Exception) -> bool:
-        if self.settings.llm_backend == LLMBackend.OLLAMA:
-            from ollama import ResponseError
-
-            return isinstance(exc, ResponseError)
-
-        if self.settings.llm_backend == LLMBackend.OPENAI_LIKE:
-            from openai import APIStatusError
-
-            return isinstance(exc, APIStatusError)
-
-        return False
-
     def _is_openai_timeout(self, exc: Exception) -> bool:
         if self.settings.llm_backend != LLMBackend.OPENAI_LIKE:
             return False

src/paperless_ai/exceptions.py

@@ -1,6 +1,2 @@
 class LLMTimeoutError(Exception):
     pass
-
-
-class LLMProviderError(Exception):
-    """The LLM backend rejected the request."""

src/paperless_ai/indexing.py

@@ -721,3 +721,20 @@ def retrieve_similar_nodes(
             continue
         filtered.append(node)
     return filtered
+
+
+def _node_document_ids(nodes: list["NodeWithScore"]) -> list[int]:
+    document_ids: list[int] = []
+    for node in nodes:
+        document_id = node.metadata.get("document_id")
+        if document_id is None:  # pragma: no cover
+            # See the matching guard in retrieve_similar_nodes() above.
+            continue
+        try:
+            document_ids.append(int(document_id))
+        except ValueError:  # pragma: no cover
+            logger.warning(
+                "Skipping LLM index result with invalid document_id %r.",
+                document_id,
+            )
+    return document_ids

src/paperless_ai/taxonomy.py

@@ -31,11 +31,6 @@ class TaxonomyCandidate(TypedDict):
     weight: float
 
 
-class SimilarDocument(TypedDict):
-    document_id: int
-    weight: float
-
-
 class TaxonomyCandidates(TypedDict):
     tags: list[TaxonomyCandidate]
     document_types: list[TaxonomyCandidate]
@@ -54,10 +49,10 @@ def empty_taxonomy_candidates() -> TaxonomyCandidates:
     )
 
 
-def _node_document_weights(nodes: list["NodeWithScore"]) -> list[SimilarDocument]:
+def _node_document_weights(nodes: list["NodeWithScore"]) -> dict[int, float]:
-    """Sum each node's similarity score into its document_id (a document can
+    """document_id -> that node's similarity score, summed if a document_id
-    appear via multiple chunks/nodes) and return one SimilarDocument per
+    appears more than once across the retrieved nodes (e.g. multiple chunks
-    distinct document_id."""
+    of the same source document)."""
     weights: dict[int, float] = defaultdict(float)
     for node in nodes:
         document_id = node.metadata.get("document_id")
@@ -70,14 +65,7 @@ def _node_document_weights(nodes: list["NodeWithScore"]) -> list[SimilarDocument
             weights[int(document_id)] += float(node.score or 0.0)
         except (TypeError, ValueError):  # pragma: no cover
             continue
-    return sorted(
+    return weights
-        (
-            SimilarDocument(document_id=document_id, weight=weight)
-            for document_id, weight in weights.items()
-        ),
-        key=lambda similar: similar["weight"],
-        reverse=True,
-    )
 
 
 def _visible_ranked_candidates(
@@ -113,25 +101,20 @@ def _visible_ranked_candidates(
 
 
 def build_taxonomy_candidates(
-    similar_documents: list[SimilarDocument],
+    nodes: list["NodeWithScore"],
     user: User | None,
 ) -> TaxonomyCandidates:
-    """Resolve each similar document's id to a live Document, read its
+    """Resolve each neighbour node's document_id to a live Document, read its
-    *current* tags/type/correspondent/storage_path via the ORM (never any
+    *current* tags/type/correspondent/storage_path via the ORM (never the
-    possibly-stale names an adapter's source might have cached), weight each
+    possibly-stale names cached in vector-index node metadata), weight each
-    distinct taxonomy object by aggregate similarity weight, permission-filter
+    distinct taxonomy object by aggregate neighbour similarity, permission-filter
     against what ``user`` can see, and return each category ranked by weight
-    and capped. ``similar_documents`` may come from either the vector-RAG
+    and capped.
-    adapter or the full-text fallback adapter - both produce this same shape.
     """
-    if not similar_documents:
-        return empty_taxonomy_candidates()
 
-    # Both adapters guarantee at most one SimilarDocument per document_id, so
+    document_weights = _node_document_weights(nodes)
-    # this never silently drops a duplicate's weight.
+    if not document_weights:
-    document_weights: dict[int, float] = {
+        return empty_taxonomy_candidates()
-        s["document_id"]: s["weight"] for s in similar_documents
-    }
 
     # Only .tags.all() needs prefetching (a reverse M2M, one extra query for
     # the whole batch). document_type/correspondent/storage_path are read

src/paperless_ai/tests/test_ai_classifier.py

@@ -1,5 +1,4 @@
 import datetime
-from collections.abc import Generator
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 from unittest.mock import patch
@@ -7,24 +6,18 @@ from unittest.mock import patch
 import pytest
 import pytest_mock
 from django.test import override_settings
-from guardian.shortcuts import assign_perm
-from guardian.shortcuts import remove_perm
 
 from documents.models import Document
-from documents.search import TantivyBackend
 from documents.tests.factories import DocumentFactory
 from documents.tests.factories import TagFactory
 from documents.tests.factories import UserFactory
 from paperless.config import AIConfig
-from paperless_ai.ai_classifier import TAXONOMY_CANDIDATE_TOP_K
-from paperless_ai.ai_classifier import _fulltext_similar_documents
 from paperless_ai.ai_classifier import build_localization_prompt
 from paperless_ai.ai_classifier import build_prompt_with_rag
 from paperless_ai.ai_classifier import build_prompt_without_rag
 from paperless_ai.ai_classifier import get_ai_document_classification
 from paperless_ai.ai_classifier import get_language_name
 from paperless_ai.ai_classifier import get_taxonomy_context
-from paperless_ai.taxonomy import SimilarDocument
 from paperless_ai.taxonomy import TaxonomyCandidate
 from paperless_ai.taxonomy import TaxonomyCandidates
 
@@ -227,10 +220,12 @@ def test_use_rag_if_configured(
 
 @pytest.mark.django_db
 @patch("paperless_ai.client.AIClient.run_llm_query")
-@patch("paperless_ai.ai_classifier.build_prompt_with_rag")
+@patch("paperless_ai.ai_classifier.build_prompt_without_rag")
+@patch("paperless_ai.ai_classifier.AIConfig")
 @override_settings(LLM_BACKEND="ollama", LLM_MODEL="some_model")
-def test_use_rag_prompt_even_without_embedding_backend(
+def test_use_without_rag_if_not_configured(
-    mock_build_prompt_with_rag,
+    mock_ai_config,
+    mock_build_prompt_without_rag,
     mock_run_llm_query,
     mock_document,
 ):
@@ -240,13 +235,13 @@ def test_use_rag_prompt_even_without_embedding_backend(
     WHEN:
         - get_ai_document_classification() is called
     THEN:
-        - The RAG-context prompt builder is still used (fed by the full-text
+        - The non-RAG prompt builder is used
-          fallback's context/candidates instead of the vector store's)
     """
-    mock_build_prompt_with_rag.return_value = "Prompt with RAG"
+    mock_ai_config.return_value.llm_embedding_backend = None
+    mock_build_prompt_without_rag.return_value = "Prompt without RAG"
     mock_run_llm_query.return_value = NESTED_SUGGESTIONS
     get_ai_document_classification(mock_document)
-    mock_build_prompt_with_rag.assert_called_once()
+    mock_build_prompt_without_rag.assert_called_once()
 
 
 @pytest.mark.django_db
@@ -325,7 +320,6 @@ def test_build_localization_prompt_preserves_unicode_characters():
 
 
 @pytest.mark.django_db
-@override_settings(LLM_EMBEDDING_BACKEND="huggingface")
 def test_get_taxonomy_context_assembles_rag_text_and_candidates():
     """
     GIVEN:
@@ -360,7 +354,6 @@ def test_get_taxonomy_context_assembles_rag_text_and_candidates():
 
 
 @pytest.mark.django_db
-@override_settings(LLM_EMBEDDING_BACKEND="huggingface")
 def test_get_taxonomy_context_preserves_similarity_order_and_distinct_documents():
     """
     GIVEN:
@@ -431,7 +424,6 @@ def test_get_taxonomy_context_preserves_similarity_order_and_distinct_documents(
 
 
 @pytest.mark.django_db
-@override_settings(LLM_EMBEDDING_BACKEND="huggingface")
 def test_get_taxonomy_context_no_similar_docs():
     """
     GIVEN:
@@ -455,67 +447,6 @@ def test_get_taxonomy_context_no_similar_docs():
     }
 
 
-@pytest.mark.django_db
-def test_get_taxonomy_context_uses_fulltext_fallback_when_no_embedding_backend(
-    mocker: pytest_mock.MockerFixture,
-) -> None:
-    """
-    GIVEN:
-        - No LLM embedding backend is configured (the default test settings)
-    WHEN:
-        - get_taxonomy_context() is called
-    THEN:
-        - _fulltext_similar_documents() is called with the document, the user
-          and TAXONOMY_CANDIDATE_TOP_K
-        - retrieve_similar_nodes() (the vector path) is never called
-    """
-    document = DocumentFactory.create(content="Some content")
-    mock_fulltext = mocker.patch(
-        "paperless_ai.ai_classifier._fulltext_similar_documents",
-        return_value=[],
-    )
-    mock_retrieve = mocker.patch("paperless_ai.ai_classifier.retrieve_similar_nodes")
-
-    get_taxonomy_context(document, user=None)
-
-    mock_fulltext.assert_called_once_with(
-        document,
-        None,
-        top_k=TAXONOMY_CANDIDATE_TOP_K,
-    )
-    mock_retrieve.assert_not_called()
-
-
-@pytest.mark.django_db
-@override_settings(LLM_EMBEDDING_BACKEND="huggingface")
-def test_get_taxonomy_context_uses_vector_path_when_embedding_backend_configured(
-    mocker: pytest_mock.MockerFixture,
-) -> None:
-    """
-    GIVEN:
-        - An LLM embedding backend is configured
-    WHEN:
-        - get_taxonomy_context() is called
-    THEN:
-        - retrieve_similar_nodes() (the vector path) is called
-        - _fulltext_similar_documents() (the no-embedding-backend fallback)
-          is never called
-    """
-    document = DocumentFactory.create(content="Some content")
-    mock_retrieve = mocker.patch(
-        "paperless_ai.ai_classifier.retrieve_similar_nodes",
-        return_value=[],
-    )
-    mock_fulltext = mocker.patch(
-        "paperless_ai.ai_classifier._fulltext_similar_documents",
-    )
-
-    get_taxonomy_context(document, user=None)
-
-    mock_retrieve.assert_called_once()
-    mock_fulltext.assert_not_called()
-
-
 class TestGetTaxonomyContextVisibility:
     """get_taxonomy_context must not materialize every visible document id
     for a user who can already see the whole library: a superuser (like no
@@ -528,7 +459,6 @@ class TestGetTaxonomyContextVisibility:
     """
 
     @pytest.mark.django_db
-    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
     def test_skips_permission_lookup_for_superuser(
         self,
         mocker: pytest_mock.MockerFixture,
@@ -547,18 +477,17 @@ class TestGetTaxonomyContextVisibility:
             "paperless_ai.ai_classifier.retrieve_similar_nodes",
             return_value=[],
         )
-        mock_permitted = mocker.patch(
+        mock_get_objects = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_object_ids",
+            "paperless_ai.ai_classifier.get_objects_for_user_owner_aware",
         )
         user = UserFactory.create(is_superuser=True)
 
         get_taxonomy_context(document, user)
 
-        mock_permitted.assert_not_called()
+        mock_get_objects.assert_not_called()
         assert mock_retrieve.call_args.kwargs["document_ids"] is None
 
     @pytest.mark.django_db
-    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
     def test_skips_permission_lookup_when_no_user(
         self,
         mocker: pytest_mock.MockerFixture,
@@ -577,17 +506,16 @@ class TestGetTaxonomyContextVisibility:
             "paperless_ai.ai_classifier.retrieve_similar_nodes",
             return_value=[],
         )
-        mock_permitted = mocker.patch(
+        mock_get_objects = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_object_ids",
+            "paperless_ai.ai_classifier.get_objects_for_user_owner_aware",
         )
 
         get_taxonomy_context(document, None)
 
-        mock_permitted.assert_not_called()
+        mock_get_objects.assert_not_called()
         assert mock_retrieve.call_args.kwargs["document_ids"] is None
 
     @pytest.mark.django_db
-    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
     def test_restricts_to_visible_documents_for_non_superuser(
         self,
         mocker: pytest_mock.MockerFixture,
@@ -598,7 +526,7 @@ class TestGetTaxonomyContextVisibility:
         WHEN:
             - get_taxonomy_context() is called
         THEN:
-            - The user's permitted document ids are looked up and passed to
+            - The user's visible document ids are looked up and passed to
               retrieve_similar_nodes() as a restriction
         """
         document = DocumentFactory.create(content="Some content")
@@ -606,232 +534,21 @@ class TestGetTaxonomyContextVisibility:
             "paperless_ai.ai_classifier.retrieve_similar_nodes",
             return_value=[],
         )
-        mock_permitted = mocker.patch(
+        mock_queryset = mocker.MagicMock()
-            "paperless_ai.ai_classifier.permitted_object_ids",
+        mock_queryset.values_list.return_value = [1, 2, 3]
-            return_value=[1, 2, 3],
+        mock_get_objects = mocker.patch(
+            "paperless_ai.ai_classifier.get_objects_for_user_owner_aware",
+            return_value=mock_queryset,
         )
         user = UserFactory.create(is_superuser=False)
 
         get_taxonomy_context(document, user)
 
-        mock_permitted.assert_called_once_with(user, Document, "view_document")
+        mock_get_objects.assert_called_once_with(user, "view_document", Document)
         assert mock_retrieve.call_args.kwargs["document_ids"] == [1, 2, 3]
 
 
 @pytest.mark.django_db
-class TestFulltextSimilarDocuments:
-    """_fulltext_similar_documents is the no-embedding-backend fallback: it
-    asks the Tantivy full-text index for "More Like This" neighbours instead
-    of the vector store, and synthesizes a rank-based weight since Tantivy's
-    more_like_this_ids returns only an ordered id list, no scores.
-    """
-
-    @pytest.fixture
-    def fulltext_backend(
-        self,
-        mocker: pytest_mock.MockerFixture,
-    ) -> Generator[TantivyBackend, None, None]:
-        """An in-memory Tantivy backend, wired up as the module-level
-        singleton _fulltext_similar_documents resolves via get_backend()."""
-        backend = TantivyBackend(path=None)
-        backend.open()
-        mocker.patch("documents.search.get_backend", return_value=backend)
-        try:
-            yield backend
-        finally:
-            backend.close()
-
-    def test_ranks_by_rank_based_weight_descending(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A source document and two similar documents indexed in Tantivy
-        WHEN:
-            - _fulltext_similar_documents() is called
-        THEN:
-            - Each result's weight reflects its rank (first result weighted
-              higher than the second), not a raw similarity score
-        """
-        source = DocumentFactory.create(content="quarterly financial report details")
-        first = DocumentFactory.create(content="quarterly financial report details")
-        second = DocumentFactory.create(content="financial report")
-        for doc in (source, first, second):
-            fulltext_backend.add_or_update(doc)
-
-        result = _fulltext_similar_documents(source, user=None, top_k=5)
-
-        assert len(result) == 2
-        weight_by_id = {s["document_id"]: s["weight"] for s in result}
-        assert weight_by_id[first.pk] > weight_by_id[second.pk]
-
-    def test_excludes_source_document(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A source document indexed in Tantivy with no other documents
-        WHEN:
-            - _fulltext_similar_documents() is called
-        THEN:
-            - An empty list is returned - the source document is never its
-              own similar document
-        """
-        source = DocumentFactory.create(content="unique unrelated content")
-        fulltext_backend.add_or_update(source)
-
-        result = _fulltext_similar_documents(source, user=None, top_k=5)
-
-        assert result == []
-
-    def test_empty_index_returns_empty_list(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A document that has never been indexed (fresh/empty Tantivy index)
-        WHEN:
-            - _fulltext_similar_documents() is called
-        THEN:
-            - An empty list is returned rather than raising
-        """
-        source = DocumentFactory.create(content="never indexed")
-
-        result = _fulltext_similar_documents(source, user=None, top_k=5)
-
-        assert result == []
-
-    def test_respects_top_k_limit(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A source document and four similar documents indexed
-        WHEN:
-            - _fulltext_similar_documents() is called with top_k=2
-        THEN:
-            - At most 2 results are returned
-        """
-        source = DocumentFactory.create(content="shared overlapping keyword text")
-        fulltext_backend.add_or_update(source)
-        for _ in range(4):
-            fulltext_backend.add_or_update(
-                DocumentFactory.create(content="shared overlapping keyword text"),
-            )
-
-        result = _fulltext_similar_documents(source, user=None, top_k=2)
-
-        assert len(result) == 2
-
-    def test_result_shape_is_similar_document(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A source document and one similar document indexed
-        WHEN:
-            - _fulltext_similar_documents() is called
-        THEN:
-            - Each result is a SimilarDocument (document_id + weight only)
-        """
-        source = DocumentFactory.create(content="shared content phrase")
-        other = DocumentFactory.create(content="shared content phrase")
-        fulltext_backend.add_or_update(source)
-        fulltext_backend.add_or_update(other)
-
-        result = _fulltext_similar_documents(source, user=None, top_k=5)
-
-        # rank 0 (the only/best result) with top_k=5 -> weight = top_k - rank = 5.0,
-        # per the "first result gets top_k, the last gets 1" formula.
-        assert result == [SimilarDocument(document_id=other.pk, weight=5.0)]
-
-    def test_superuser_sees_other_users_documents(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A source document owned by one user and a similar document
-              owned by a different user, with no sharing between them
-        WHEN:
-            - _fulltext_similar_documents() is called with a superuser
-        THEN:
-            - The other user's document is still returned as a similar
-              document - a superuser must not be narrowed by the backend's
-              owner-based permission filter
-        """
-        owner = UserFactory.create()
-        other_owner = UserFactory.create()
-        superuser = UserFactory.create(is_superuser=True)
-        source = DocumentFactory.create(
-            content="shared content phrase",
-            owner=owner,
-        )
-        other = DocumentFactory.create(
-            content="shared content phrase",
-            owner=other_owner,
-        )
-        fulltext_backend.add_or_update(source)
-        fulltext_backend.add_or_update(other)
-
-        result = _fulltext_similar_documents(source, user=superuser, top_k=5)
-
-        assert [s["document_id"] for s in result] == [other.pk]
-
-    def test_excludes_stale_permitted_document_for_regular_user(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A regular (non-superuser) user
-            - A similar document the user is permitted to view, and another
-              similar document indexed while the user still had view
-              permission but which has since had that permission revoked in
-              the database, i.e. the Tantivy index has stale permission data
-        WHEN:
-            - _fulltext_similar_documents() is called with that user
-        THEN:
-            - Only the still-permitted document is returned - the DB
-              re-check via restrict_queryset_to_visible() must catch the
-              document Tantivy's stale index still thinks is visible
-        """
-        owner = UserFactory.create()
-        viewer = UserFactory.create(is_superuser=False)
-        source = DocumentFactory.create(
-            content="shared content phrase",
-            owner=owner,
-        )
-        permitted = DocumentFactory.create(
-            content="shared content phrase",
-            owner=owner,
-        )
-        now_private = DocumentFactory.create(
-            content="shared content phrase",
-            owner=owner,
-        )
-        assign_perm("view_document", viewer, permitted)
-        assign_perm("view_document", viewer, now_private)
-        fulltext_backend.add_or_update(source)
-        fulltext_backend.add_or_update(permitted)
-        fulltext_backend.add_or_update(now_private)
-
-        # Revoke access after indexing, without reindexing: the index still
-        # carries viewer as a permitted viewer for `now_private`.
-        remove_perm("view_document", viewer, now_private)
-
-        result = _fulltext_similar_documents(source, user=viewer, top_k=5)
-
-        assert [s["document_id"] for s in result] == [permitted.pk]
-
-
-@pytest.mark.django_db
-@override_settings(LLM_EMBEDDING_BACKEND="huggingface")
 @patch("paperless_ai.ai_classifier.retrieve_similar_nodes")
 def test_get_taxonomy_context_retrieval_failure_degrades_to_no_hints(mock_retrieve):
     """
@@ -858,7 +575,6 @@ def test_get_taxonomy_context_retrieval_failure_degrades_to_no_hints(mock_retrie
 
 
 @pytest.mark.django_db
-@override_settings(LLM_EMBEDDING_BACKEND="huggingface")
 @patch("paperless_ai.ai_classifier.build_taxonomy_candidates")
 @patch("paperless_ai.ai_classifier.retrieve_similar_nodes")
 def test_get_taxonomy_context_candidate_building_failure_degrades_to_no_hints(

src/paperless_ai/tests/test_ai_indexing.py

@@ -1188,7 +1188,9 @@ class TestRetrieveSimilarNodesAgainstRealIndex:
 
         nodes = indexing.retrieve_similar_nodes(a, document_ids=[b.id])
 
-        assert all(int(node.metadata["document_id"]) == b.id for node in nodes)
+        assert all(
+            document_id == b.id for document_id in indexing._node_document_ids(nodes)
+        )
 
     def test_excludes_self(
         self,
@@ -1210,7 +1212,7 @@ class TestRetrieveSimilarNodesAgainstRealIndex:
 
         nodes = indexing.retrieve_similar_nodes(a, top_k=5)
 
-        assert {int(node.metadata["document_id"]) for node in nodes} == {b.id}
+        assert set(indexing._node_document_ids(nodes)) == {b.id}
 
     def test_excludes_self_with_multiple_chunks(
         self,
@@ -1233,4 +1235,4 @@ class TestRetrieveSimilarNodesAgainstRealIndex:
 
         nodes = indexing.retrieve_similar_nodes(a, top_k=3)
 
-        assert {int(node.metadata["document_id"]) for node in nodes} == {b.id}
+        assert set(indexing._node_document_ids(nodes)) == {b.id}

src/paperless_ai/tests/test_client.py

@@ -4,7 +4,6 @@ from unittest.mock import MagicMock
 from unittest.mock import patch
 
 import httpx
-import ollama
 import openai
 import pytest
 from llama_index.core.llms.llm import ToolSelection
@@ -12,7 +11,6 @@ from llama_index.core.llms.llm import ToolSelection
 from paperless_ai.client import LLM_SYSTEM_PROMPT
 from paperless_ai.client import PLACEHOLDER_API_KEY
 from paperless_ai.client import AIClient
-from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 
 
@@ -216,52 +214,6 @@ def test_run_llm_query_openai_timeout_raises_local_error(
         client.run_llm_query("test_prompt")
 
 
-def test_run_llm_query_openai_status_error_raises_provider_error(
-    mock_ai_config,
-    mock_openai_llm,
-):
-    mock_ai_config.llm_backend = "openai-like"
-    mock_ai_config.llm_model = "test_model"
-    mock_ai_config.llm_endpoint = "http://test-url"
-
-    request = httpx.Request("POST", "http://test-url/v1/chat/completions")
-    body = {"error": {"message": "Thinking mode does not support this tool_choice"}}
-    mock_openai_llm.return_value.chat_with_tools.side_effect = openai.BadRequestError(
-        "Error code: 400",
-        response=httpx.Response(400, request=request, json=body),
-        body=body,
-    )
-
-    client = AIClient()
-
-    with pytest.raises(LLMProviderError) as exc_info:
-        client.run_llm_query("test_prompt")
-    assert str(exc_info.value) == ""
-    assert isinstance(exc_info.value.__cause__, openai.BadRequestError)
-
-
-def test_run_llm_query_ollama_response_error_raises_provider_error(
-    mock_ai_config,
-    mock_ollama_llm,
-):
-    mock_ai_config.llm_backend = "ollama"
-    mock_ai_config.llm_model = "test_model"
-    mock_ai_config.llm_endpoint = "http://test-url"
-
-    response_error = ollama.ResponseError(
-        "confidential provider response",
-        status_code=400,
-    )
-    mock_ollama_llm.return_value.chat.side_effect = response_error
-
-    client = AIClient()
-
-    with pytest.raises(LLMProviderError) as exc_info:
-        client.run_llm_query("test_prompt")
-    assert str(exc_info.value) == ""
-    assert exc_info.value.__cause__ is response_error
-
-
 def test_run_llm_query_httpx_timeout_raises_local_error(
     mock_ai_config,
     mock_ollama_llm,

src/paperless_ai/tests/test_taxonomy.py

@@ -1,4 +1,5 @@
 import json
+from types import SimpleNamespace
 
 import pytest
 import pytest_mock
@@ -9,14 +10,14 @@ from documents.tests.factories import DocumentTypeFactory
 from documents.tests.factories import StoragePathFactory
 from documents.tests.factories import TagFactory
 from documents.tests.factories import UserFactory
-from paperless_ai.taxonomy import SimilarDocument
 from paperless_ai.taxonomy import TaxonomyCandidates
 from paperless_ai.taxonomy import build_taxonomy_candidates
 from paperless_ai.taxonomy import format_taxonomy_for_prompt
 
 
-def make_similar(document_id: int, weight: float) -> SimilarDocument:
+def make_node(document_id: int, score: float) -> SimpleNamespace:
-    return SimilarDocument(document_id=document_id, weight=weight)
+    """A stand-in for NodeWithScore: only ``.metadata``/``.score`` are read."""
+    return SimpleNamespace(metadata={"document_id": str(document_id)}, score=score)
 
 
 @pytest.mark.django_db
@@ -52,9 +53,9 @@ class TestBuildTaxonomyCandidates:
         doc_a.tags.add(tag)
         doc_b = DocumentFactory.create()
         doc_b.tags.add(tag)
-        similar_documents = [make_similar(doc_a.pk, 0.9), make_similar(doc_b.pk, 0.4)]
+        nodes = [make_node(doc_a.pk, 0.9), make_node(doc_b.pk, 0.4)]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["tags"]) == 1
         assert result["tags"][0]["id"] == tag.pk
@@ -79,9 +80,9 @@ class TestBuildTaxonomyCandidates:
         document.tags.add(tag)
         tag.name = "New Name"
         tag.save()
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert result["tags"][0]["name"] == "New Name"
 
@@ -101,9 +102,9 @@ class TestBuildTaxonomyCandidates:
         document = DocumentFactory.create()
         document.tags.add(tag)
         tag.delete()
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert result["tags"] == []
 
@@ -122,12 +123,9 @@ class TestBuildTaxonomyCandidates:
         strong_doc.tags.add(strong_tag)
         weak_doc = DocumentFactory.create()
         weak_doc.tags.add(weak_tag)
-        similar_documents = [
+        nodes = [make_node(strong_doc.pk, 0.9), make_node(weak_doc.pk, 0.1)]
-            make_similar(strong_doc.pk, 0.9),
-            make_similar(weak_doc.pk, 0.1),
-        ]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert [c["name"] for c in result["tags"]] == ["Strong", "Weak"]
 
@@ -143,9 +141,9 @@ class TestBuildTaxonomyCandidates:
         document = DocumentFactory.create()
         for i in range(15):
             document.tags.add(TagFactory.create(name=f"Tag{i}"))
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["tags"]) == 10
 
@@ -159,12 +157,12 @@ class TestBuildTaxonomyCandidates:
             - Only 5 correspondents are returned
         """
         correspondents = CorrespondentFactory.create_batch(7)
-        similar_documents = [
+        nodes = [
-            make_similar(DocumentFactory.create(correspondent=c).pk, 0.5)
+            make_node(DocumentFactory.create(correspondent=c).pk, 0.5)
             for c in correspondents
         ]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["correspondents"]) == 5
 
@@ -179,9 +177,9 @@ class TestBuildTaxonomyCandidates:
         """
         document_type = DocumentTypeFactory.create(name="Invoice")
         document = DocumentFactory.create(document_type=document_type)
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["document_types"]) == 1
         assert result["document_types"][0]["id"] == document_type.pk
@@ -197,12 +195,12 @@ class TestBuildTaxonomyCandidates:
             - Only 5 document_types are returned
         """
         document_types = DocumentTypeFactory.create_batch(7)
-        similar_documents = [
+        nodes = [
-            make_similar(DocumentFactory.create(document_type=dt).pk, 0.5)
+            make_node(DocumentFactory.create(document_type=dt).pk, 0.5)
             for dt in document_types
         ]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["document_types"]) == 5
 
@@ -217,9 +215,9 @@ class TestBuildTaxonomyCandidates:
         """
         storage_path = StoragePathFactory.create(name="Invoices")
         document = DocumentFactory.create(storage_path=storage_path)
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["storage_paths"]) == 1
         assert result["storage_paths"][0]["id"] == storage_path.pk
@@ -235,12 +233,12 @@ class TestBuildTaxonomyCandidates:
             - Only 5 storage_paths are returned
         """
         storage_paths = StoragePathFactory.create_batch(7)
-        similar_documents = [
+        nodes = [
-            make_similar(DocumentFactory.create(storage_path=sp).pk, 0.5)
+            make_node(DocumentFactory.create(storage_path=sp).pk, 0.5)
             for sp in storage_paths
         ]
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert len(result["storage_paths"]) == 5
 
@@ -260,14 +258,14 @@ class TestBuildTaxonomyCandidates:
         tag = TagFactory.create(name="Restricted")
         document = DocumentFactory.create()
         document.tags.add(tag)
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
         user = UserFactory.create()
         mocker.patch(
             "documents.permissions.permitted_object_ids",
             return_value=[],  # user cannot see this tag
         )
 
-        result = build_taxonomy_candidates(similar_documents, user=user)
+        result = build_taxonomy_candidates(nodes, user=user)
 
         assert result["tags"] == []
 
@@ -297,10 +295,10 @@ class TestBuildTaxonomyCandidates:
         tag.save()
         document = DocumentFactory.create()
         document.tags.add(tag)
-        similar_documents = [make_similar(document.pk, 0.5)]
+        nodes = [make_node(document.pk, 0.5)]
         spy = mocker.patch("documents.permissions.permitted_object_ids")
 
-        result = build_taxonomy_candidates(similar_documents, user=None)
+        result = build_taxonomy_candidates(nodes, user=None)
 
         assert result["tags"][0]["name"] == "Owned"
         spy.assert_not_called()

src/paperless_mail/mail.py

@@ -6,7 +6,6 @@ import socket
 import ssl
 import tempfile
 import traceback
-import unicodedata
 from datetime import date
 from datetime import timedelta
 from fnmatch import fnmatch
@@ -45,6 +44,7 @@ from documents.models import Correspondent
 from documents.models import PaperlessTask
 from documents.parsers import is_mime_type_supported
 from documents.tasks import consume_file
+from documents.utils import normalize_unicode
 from paperless.network import is_public_ip
 from paperless.network import resolve_hostname_ips
 from paperless_mail.models import MailAccount
@@ -617,10 +617,10 @@ class MailAccountHandler(LoggingMixin):
         rule: MailRule,
     ) -> str | None:
         if rule.assign_title_from == MailRule.TitleSource.FROM_SUBJECT:
-            return unicodedata.normalize("NFC", message.subject)
+            return normalize_unicode(message.subject)
 
         elif rule.assign_title_from == MailRule.TitleSource.FROM_FILENAME:
-            return unicodedata.normalize("NFC", Path(att.filename).stem)
+            return normalize_unicode(Path(att.filename).stem)
 
         elif rule.assign_title_from == MailRule.TitleSource.NONE:
             return None
@@ -1004,6 +1004,8 @@ class MailAccountHandler(LoggingMixin):
         consume_tasks = []
 
         for att in message.attachments:
+            attachment_filename = normalize_unicode(att.filename)
+
             if (
                 att.content_disposition != "attachment"
                 and rule.attachment_type
@@ -1018,7 +1020,7 @@ class MailAccountHandler(LoggingMixin):
 
             if not self.filename_inclusion_matches(
                 rule.filter_attachment_filename_include,
-                att.filename,
+                attachment_filename,
             ):
                 # Force the filename and pattern to the lowercase
                 # as this is system dependent otherwise
@@ -1030,7 +1032,7 @@ class MailAccountHandler(LoggingMixin):
                 continue
             elif self.filename_exclusion_matches(
                 rule.filter_attachment_filename_exclude,
-                att.filename,
+                attachment_filename,
             ):
                 self.log.debug(
                     f"Rule {rule}: "
@@ -1064,7 +1066,7 @@ class MailAccountHandler(LoggingMixin):
                 )
 
                 attachment_name = pathvalidate.sanitize_filename(
-                    unicodedata.normalize("NFC", att.filename),
+                    attachment_filename,
                 )
                 if attachment_name:
                     temp_filename = temp_dir / attachment_name
@@ -1175,7 +1177,7 @@ class MailAccountHandler(LoggingMixin):
         doc_overrides = DocumentMetadataOverrides(
             title=message.subject,
             filename=pathvalidate.sanitize_filename(
-                unicodedata.normalize("NFC", f"{message.subject}.eml"),
+                normalize_unicode(f"{message.subject}.eml"),
             ),
             correspondent_id=correspondent.id if correspondent else None,
             document_type_id=doc_type.id if doc_type else None,

src/paperless_mail/serialisers.py

@@ -8,6 +8,7 @@ from documents.serialisers import CorrespondentField
 from documents.serialisers import DocumentTypeField
 from documents.serialisers import OwnedObjectSerializer
 from documents.serialisers import TagsField
+from documents.utils import normalize_unicode
 from paperless_mail.models import MailAccount
 from paperless_mail.models import MailRule
 from paperless_mail.models import ProcessedMail
@@ -161,6 +162,12 @@ class MailRuleSerializer(OwnedObjectSerializer):
             raise serializers.ValidationError("Maximum mail age is unreasonably large.")
         return value
 
+    def validate_filter_attachment_filename_include(self, value):
+        return normalize_unicode(value)
+
+    def validate_filter_attachment_filename_exclude(self, value):
+        return normalize_unicode(value)
+
 
 class ProcessedMailSerializer(OwnedObjectSerializer):
     class Meta:

src/paperless_mail/tasks.py

@@ -1,9 +1,7 @@
 import logging
 
-from celery import Task
 from celery import shared_task
 
-from documents.models import PaperlessTask
 from paperless_mail.mail import MailAccountHandler
 from paperless_mail.mail import MailError
 from paperless_mail.models import MailAccount
@@ -12,26 +10,8 @@ from paperless_mail.models import MailRule
 logger = logging.getLogger("paperless.mail.tasks")
 
 
-@shared_task(bind=True)
+@shared_task
-def process_mail_accounts(self: Task, account_ids: list[int] | None = None) -> str:
+def process_mail_accounts(account_ids: list[int] | None = None) -> str:
-    # A scheduled check can still be running (or queued) when the next one
-    # ProcessedMail dedup only records a message once its
-    # handling has finished, so an overlapping run can still pick up the same
-    # not-yet-recorded message. Skip outright rather than race it.
-    other_mail_fetch_running = (
-        PaperlessTask.objects.filter(
-            task_type=PaperlessTask.TaskType.MAIL_FETCH,
-            status__in=[PaperlessTask.Status.PENDING, PaperlessTask.Status.STARTED],
-        )
-        .exclude(task_id=self.request.id)
-        .exists()
-    )
-    if other_mail_fetch_running:
-        logger.info(
-            "Mail account processing is already running; skipping this run.",
-        )
-        return "Skipped: mail account processing already in progress."
-
     total_new_documents = 0
     accounts = (
         MailAccount.objects.filter(pk__in=account_ids)

src/paperless_mail/tests/test_process_mail_accounts_overlap.py

@@ -1,134 +0,0 @@
-from typing import Final
-
-import pytest
-import pytest_mock
-
-from documents.models import PaperlessTask
-from documents.tests.factories import PaperlessTaskFactory
-from paperless_mail import tasks
-from paperless_mail.tests.factories import MailAccountFactory
-from paperless_mail.tests.factories import MailRuleFactory
-
-NO_DOCUMENTS_ADDED: Final = "No new documents were added."
-SKIPPED: Final = "Skipped: mail account processing already in progress."
-
-
-@pytest.mark.django_db
-@pytest.mark.usefixtures("account_with_rule")
-class TestProcessMailAccountsOverlap:
-    @pytest.fixture
-    def account_with_rule(self) -> None:
-        """An enabled mail account with a single enabled rule."""
-        account = MailAccountFactory.create()
-        MailRuleFactory.create(account=account, enabled=True)
-
-    @pytest.mark.parametrize(
-        ("status", "expected_result", "expected_call_count"),
-        [
-            pytest.param(
-                PaperlessTask.Status.PENDING,
-                SKIPPED,
-                0,
-                id="pending-task-blocks",
-            ),
-            pytest.param(
-                PaperlessTask.Status.STARTED,
-                SKIPPED,
-                0,
-                id="started-task-blocks",
-            ),
-            pytest.param(
-                PaperlessTask.Status.SUCCESS,
-                NO_DOCUMENTS_ADDED,
-                1,
-                id="finished-task-does-not-block",
-            ),
-        ],
-    )
-    def test_skips_only_while_another_mail_fetch_task_runs(
-        self,
-        mocker: pytest_mock.MockerFixture,
-        status: PaperlessTask.Status,
-        expected_result: str,
-        expected_call_count: int,
-    ) -> None:
-        """
-        GIVEN:
-            - An enabled mail account with a rule
-            - Another mail fetch task row in the given status
-        WHEN:
-            - Mail accounts are processed
-        THEN:
-            - Processing is skipped only if that other task is pending or running
-        """
-        PaperlessTaskFactory.create(
-            task_type=PaperlessTask.TaskType.MAIL_FETCH,
-            trigger_source=PaperlessTask.TriggerSource.SCHEDULED,
-            status=status,
-        )
-
-        mocked_handle = mocker.patch.object(
-            tasks.MailAccountHandler,
-            "handle_mail_account",
-            return_value=0,
-        )
-
-        result = tasks.process_mail_accounts()
-
-        assert mocked_handle.call_count == expected_call_count
-        assert result == expected_result
-
-    def test_runs_when_no_other_mail_fetch_task_exists(
-        self,
-        mocker: pytest_mock.MockerFixture,
-    ) -> None:
-        """
-        GIVEN:
-            - An enabled mail account with a rule
-            - No other mail fetch task rows
-        WHEN:
-            - Mail accounts are processed
-        THEN:
-            - The account is handled
-        """
-        mocked_handle = mocker.patch.object(
-            tasks.MailAccountHandler,
-            "handle_mail_account",
-            return_value=0,
-        )
-
-        result = tasks.process_mail_accounts()
-
-        mocked_handle.assert_called_once()
-        assert result == NO_DOCUMENTS_ADDED
-
-    def test_does_not_skip_due_to_its_own_task_row(
-        self,
-        mocker: pytest_mock.MockerFixture,
-    ) -> None:
-        """
-        GIVEN:
-            - An enabled mail account with a rule
-            - A running mail fetch task row belonging to this very task
-        WHEN:
-            - Mail accounts are processed under that task id
-        THEN:
-            - The task does not skip itself and handles the account
-        """
-        PaperlessTaskFactory.create(
-            task_id="self-task-id",
-            task_type=PaperlessTask.TaskType.MAIL_FETCH,
-            trigger_source=PaperlessTask.TriggerSource.SCHEDULED,
-            status=PaperlessTask.Status.STARTED,
-        )
-
-        mocked_handle = mocker.patch.object(
-            tasks.MailAccountHandler,
-            "handle_mail_account",
-            return_value=0,
-        )
-
-        result = tasks.process_mail_accounts.apply(task_id="self-task-id").result
-
-        mocked_handle.assert_called_once()
-        assert result == NO_DOCUMENTS_ADDED