Fix: index document versions as their root document (#14422)

* Fix: index document versions as their root document

The search index and the LLM index hold root documents only, a root being
indexed with its newest version's content. Every write path therefore had to
remember to hand them the root. Several did not.

WriteBatch.add_or_update, WriteBatch.add_or_update_ids and
llm_index_add_or_update_document now resolve a version to its root themselves.
The reindex command and update_llm_index only walk root documents, and an
incremental LLM update is scoped to the roots of the given ids through
versioning.root_document_ids, which add_or_update_ids shares. more_like_id
returns the root's id for a version. The reprocess task no longer picks the root
for the indexes and only still clears the caches of both documents.

* Fix: rebuild old search indexes and look versions up as their root for similar documents

An index built before versions were indexed as their root can still hold
version entries under their own id and metadata, and only a rebuild clears
them. Bump the schema version so the index is rebuilt on upgrade.

more_like_this_ids looked the given id up in the index as is, so asking for
documents like a version, as the AI suggestions endpoint can, silently found
nothing. It now looks the root up instead.

src/documents/management/commands/document_index.py

@@ -67,18 +67,22 @@ class Command(PaperlessCommand):
                 if options.get("recreate"):
                     wipe_index(settings.INDEX_DIR)
 
-                documents = Document.objects.select_related(
-                    "correspondent",
-                    "document_type",
-                    "storage_path",
-                    "owner",
-                ).prefetch_related(
-                    "tags",
-                    "notes__user",
-                    "custom_fields__field",
-                    "versions",
-                    "barcodes",
-                    "versions__barcodes",
+                documents = (
+                    Document.objects.filter(root_document__isnull=True)
+                    .select_related(
+                        "correspondent",
+                        "document_type",
+                        "storage_path",
+                        "owner",
+                    )
+                    .prefetch_related(
+                        "tags",
+                        "notes__user",
+                        "custom_fields__field",
+                        "versions",
+                        "barcodes",
+                        "versions__barcodes",
+                    )
                 )
                 total = documents.count()
                 rebuild_kwargs = {}

src/documents/search/_backend.py

@@ -284,9 +284,14 @@ class WriteBatch:
         and adding the new version. This ensures stale document data (e.g., after
         permission changes) doesn't persist in the index.
 
+        Only root documents are indexed, with their effective content, so a
+        version is indexed as its root document.
+
         Args:
             document: Django Document instance to index
         """
+        if document.root_document_id is not None:
+            document = document.root_document
         self.remove(document.pk)
         doc = self._backend._build_tantivy_doc(document)
         self._writer.add_document(doc)
@@ -311,20 +316,22 @@ class WriteBatch:
         An id with no matching document (e.g. deleted between the caller
         collecting ids and the batch running) is silently skipped, matching
         ``add_or_update()``'s existing single-document deferred-task behavior
-        rather than erroring or leaving a stale index entry.
+        rather than erroring or leaving a stale index entry. The id of a
+        version stands for its root document.
 
         Args:
             ids: Primary keys of Document instances to index
         """
         from documents.models import Document
         from documents.versioning import annotate_effective_content
+        from documents.versioning import root_document_ids
 
         ids = list(ids)
         if not ids:
             return
 
         queryset = annotate_effective_content(
-            Document.objects.filter(pk__in=ids)
+            Document.objects.filter(pk__in=root_document_ids(ids))
             .select_related("correspondent", "document_type", "storage_path", "owner")
             .prefetch_related(
                 "tags",
@@ -1036,16 +1043,21 @@ class TantivyBackend:
         excluded from results.
 
         Args:
-            doc_id: Primary key of the reference document
+            doc_id: Primary key of the reference document, or of one of its
+                versions
             user: User for permission filtering (None for no filtering)
             limit: Maximum number of IDs to return (None = all matching docs)
 
         Returns:
             List of similar document IDs (excluding the original)
         """
+        from documents.versioning import root_document_ids
+
         self._ensure_open()
         searcher = self._index.searcher()
 
+        # Only root documents are indexed, so a version stands for its root
+        doc_id = next(iter(root_document_ids([doc_id])), doc_id)
         id_query = tantivy.Query.term_query(self._schema, "id", doc_id)
         results = searcher.search(id_query, limit=1)
 

src/documents/search/_schema.py

@@ -28,7 +28,9 @@ logger = logging.getLogger("paperless.search")
 #      columns dropped. tantivy compares schemas by ordered field list, so an
 #      index built by v1 rejects every write against the v2 schema.
 # v3 - barcodes JSON field for stored barcode contents
-SCHEMA_VERSION: Final[int] = 3
+# v4 - root documents only. Earlier indexes may hold document versions under
+#      their own id and metadata, so they are rebuilt without them.
+SCHEMA_VERSION: Final[int] = 4
 
 # Present in the index directory from the moment a full rebuild starts until it
 # finishes. If a rebuild is interrupted it is left behind, so the half-built

src/documents/tasks.py

@@ -490,23 +490,20 @@ def update_document_content_maybe_archive_file(
                     shutil.move(thumbnail, document.thumbnail_path)
 
             document.refresh_from_db()
-            root_document = (
-                document.root_document if document.root_document_id else document
-            )
             logger.info(
-                f"Updating index for document {root_document.pk} ({document.archive_checksum})",
+                f"Updating index for document {document_id} ({document.archive_checksum})",
             )
             from documents.search import get_backend
 
-            get_backend().add_or_update(root_document)
+            get_backend().add_or_update(document)
 
             ai_config = AIConfig()
             if ai_config.llm_index_enabled:
-                llm_index_add_or_update_document(root_document)
+                llm_index_add_or_update_document(document)
 
             clear_document_caches(document.pk)
-            if root_document.pk != document.pk:
-                clear_document_caches(root_document.pk)
+            if document.root_document_id is not None:
+                clear_document_caches(document.root_document_id)
 
         except Exception:
             logger.exception(

src/documents/tests/search/test_backend.py

@@ -293,6 +293,80 @@ class TestAddOrUpdateIds:
         assert backend.search_ids("updated", user=None) == [doc.pk]
 
 
+class TestVersionsAreIndexedAsTheirRoot:
+    """Only root documents are indexed, with their effective content, so
+    every write path that is handed a version indexes its root instead."""
+
+    @staticmethod
+    def _root_with_version() -> tuple[Document, Document]:
+        root = DocumentFactory(title="Statement", content="stale text")
+        version = DocumentFactory(
+            title="Statement",
+            content="latest text",
+            root_document=root,
+            version_index=1,
+        )
+        return root, version
+
+    def test_add_or_update_indexes_the_root_of_a_version(
+        self,
+        backend: TantivyBackend,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - The version is passed to add_or_update
+        THEN:
+            - The root is indexed with the version's text, and the version is not
+        """
+        root, version = self._root_with_version()
+
+        backend.add_or_update(version)
+
+        assert backend.search_ids("latest", user=None) == [root.pk]
+        assert backend.search_ids("stale", user=None) == []
+
+    def test_add_or_update_ids_indexes_each_root_once(
+        self,
+        backend: TantivyBackend,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - Both ids are passed to add_or_update_ids
+        THEN:
+            - The root is indexed once and the version is not indexed
+        """
+        root, version = self._root_with_version()
+
+        with backend.batch_update() as batch:
+            batch.add_or_update_ids([version.pk, root.pk])
+
+        assert backend.search_ids("Statement", user=None) == [root.pk]
+        assert backend.search_ids("latest", user=None) == [root.pk]
+
+    def test_add_or_update_ids_resolves_a_lone_version(
+        self,
+        backend: TantivyBackend,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - Only the version's id is passed to add_or_update_ids
+        THEN:
+            - The root is indexed
+        """
+        root, version = self._root_with_version()
+
+        with backend.batch_update() as batch:
+            batch.add_or_update_ids([version.pk])
+
+        assert backend.search_ids("latest", user=None) == [root.pk]
+
+
 class TestSearch:
     """Test search query parsing and matching via search_ids."""
 
@@ -1062,6 +1136,34 @@ class TestMoreLikeThis:
         assert 150 not in ids
         assert 151 in ids
 
+    def test_more_like_this_ids_seeded_by_version_uses_root(
+        self,
+        backend: TantivyBackend,
+    ) -> None:
+        """A version is not indexed, so it must be looked up as its root."""
+        root = DocumentFactory.create(
+            title="Important document",
+            content="financial information report",
+        )
+        version = DocumentFactory.create(
+            title="Important document",
+            content="financial information report",
+            root_document=root,
+            version_index=1,
+        )
+        other = DocumentFactory.create(
+            title="Another document",
+            content="financial information report",
+        )
+        backend.add_or_update(root)
+        backend.add_or_update(other)
+
+        ids = backend.more_like_this_ids(doc_id=version.pk, user=None)
+
+        assert other.pk in ids
+        assert root.pk not in ids
+        assert version.pk not in ids
+
 
 class TestSingleton:
     """Test get_backend() and reset_backend() singleton lifecycle."""

src/documents/tests/test_api_search.py

@@ -1196,6 +1196,59 @@ class TestDocumentSearchApi(DirectoriesMixin, APITestCase):
         self.assertIn(d3.id, result_ids)
         self.assertNotIn(d4.id, result_ids)
 
+    def test_search_more_like_version_uses_its_root(self) -> None:
+        """
+        GIVEN:
+            - A document similar in content to a root document, and one that is not
+            - A version of the root document, which is never indexed
+        WHEN:
+            - API request for more like the version
+        THEN:
+            - The documents similar to the root are returned, not the version
+        """
+        indexed = {}
+        for name, title, content, day in (
+            ("root", "bank statement 1", "things i paid for in august", (2019, 3, 4)),
+            (
+                "similar",
+                "bank statement 3",
+                "things i paid for in september",
+                (2020, 7, 9),
+            ),
+            (
+                "other",
+                "Quarterly Report",
+                "quarterly revenue profit margin",
+                (2021, 11, 30),
+            ),
+        ):
+            with time_machine.travel(
+                timezone.make_aware(datetime.datetime(*day)),
+                tick=False,
+            ):
+                indexed[name] = DocumentFactory(
+                    title=title,
+                    content=content,
+                    created=datetime.date(*day),
+                    added=timezone.make_aware(datetime.datetime(*day)),
+                )
+        version = DocumentFactory(
+            root_document=indexed["root"],
+            version_index=1,
+            content="things i paid for in august",
+        )
+        backend = get_backend()
+        for document in indexed.values():
+            backend.add_or_update(document)
+
+        response = self.client.get(f"/api/documents/?more_like_id={version.id}")
+
+        self.assertEqual(response.status_code, status.HTTP_200_OK)
+        result_ids = [r["id"] for r in response.data["results"]]
+        self.assertIn(indexed["similar"].id, result_ids)
+        self.assertNotIn(indexed["other"].id, result_ids)
+        self.assertNotIn(version.id, result_ids)
+
     def test_more_like_requires_id_of_existing_document(self) -> None:
         """
         GIVEN:

src/documents/tests/test_management.py

@@ -22,6 +22,7 @@ from documents.models import Document
 from documents.tasks import update_document_content_maybe_archive_file
 from paperless_testing.assertions import FileSystemAssertsMixin
 from paperless_testing.dirs import DirectoriesMixin
+from paperless_testing.factories import DocumentFactory
 
 sample_file: Path = Path(__file__).parent / "samples" / "simple.pdf"
 
@@ -116,6 +117,27 @@ class TestMakeIndex:
         call_command("document_index", "reindex", skip_checks=True)
         mock_get_backend.return_value.rebuild.assert_called_once()
 
+    def test_reindex_skips_versions(self, mocker: MockerFixture) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - The reindex command runs
+        THEN:
+            - Only the root document is handed to the rebuild, since a version
+              is indexed as its root
+        """
+        root = DocumentFactory()
+        DocumentFactory(root_document=root, version_index=1)
+        mock_get_backend = mocker.patch(
+            "documents.management.commands.document_index.get_backend",
+        )
+
+        call_command("document_index", "reindex", skip_checks=True)
+
+        documents = mock_get_backend.return_value.rebuild.call_args.args[0]
+        assert list(documents.values_list("pk", flat=True)) == [root.pk]
+
     def test_optimize(self) -> None:
         """Optimize command must execute without error (Tantivy handles optimization automatically)."""
         call_command("document_index", "optimize", skip_checks=True)

src/documents/tests/test_tasks.py

@@ -310,7 +310,7 @@ class TestUpdateContent(DirectoriesMixin, TestCase):
 
     @mock.patch("documents.tasks.clear_document_caches")
     @mock.patch("documents.search.get_backend")
-    def test_update_content_version_indexes_root(
+    def test_update_content_version_clears_caches_for_root(
         self,
         mock_get_backend: mock.Mock,
         mock_clear_caches: mock.Mock,
@@ -321,8 +321,8 @@ class TestUpdateContent(DirectoriesMixin, TestCase):
         WHEN:
             - Update content task is called for the version
         THEN:
-            - The version's content is updated
-            - The root document is indexed rather than the version
+            - The version's content is updated, not the root's
+            - The document is indexed
             - Caches are cleared for both
         """
         root, version = self._create_root_with_version()
@@ -334,8 +334,7 @@ class TestUpdateContent(DirectoriesMixin, TestCase):
             "my document",
         )
         self.assertEqual(Document.objects.get(pk=root.pk).content, "root content")
-        indexed = mock_get_backend.return_value.add_or_update.call_args.args[0]
-        self.assertEqual(indexed.pk, root.pk)
+        mock_get_backend.return_value.add_or_update.assert_called_once()
         mock_clear_caches.assert_has_calls(
             [mock.call(version.pk), mock.call(root.pk)],
         )
@@ -343,7 +342,7 @@ class TestUpdateContent(DirectoriesMixin, TestCase):
     @override_settings(AI_ENABLED=True, LLM_EMBEDDING_BACKEND="huggingface")
     @mock.patch("documents.tasks.llm_index_add_or_update_document")
     @mock.patch("documents.search.get_backend")
-    def test_update_content_version_updates_llm_index_for_root(
+    def test_update_content_version_updates_llm_index(
         self,
         mock_get_backend: mock.Mock,
         mock_llm_index: mock.Mock,
@@ -355,14 +354,13 @@ class TestUpdateContent(DirectoriesMixin, TestCase):
         WHEN:
             - Update content task is called for the version
         THEN:
-            - The LLM index is updated for the root document, not the version
+            - The LLM index is updated
         """
-        root, version = self._create_root_with_version()
+        _, version = self._create_root_with_version()
 
         tasks.update_document_content_maybe_archive_file(version.pk)
 
         mock_llm_index.assert_called_once()
-        self.assertEqual(mock_llm_index.call_args.args[0].pk, root.pk)
 
 
 class TestUpdateContentRemoteOCR(DirectoriesMixin, TestCase):

src/documents/versioning.py

@@ -17,9 +17,26 @@ from django.db.models.functions import RowNumber
 from documents.models import Document
 
 if TYPE_CHECKING:
+    from collections.abc import Iterable
+
     from rest_framework.request import Request
 
 
+def root_document_ids(ids: Iterable[int]) -> QuerySet[int]:
+    """
+    The ids of the root documents of the given documents: a root stands for
+    itself and a version for its root. Only the indexes' bookkeeping needs
+    this, since they hold root documents only.
+    """
+    return (
+        Document.objects.filter(pk__in=ids)
+        .annotate(root_id=Coalesce("root_document_id", "id"))
+        .order_by()
+        .values_list("root_id", flat=True)
+        .distinct()
+    )
+
+
 def versions_newest_first(documents: QuerySet[Document]) -> QuerySet[Document]:
     """
     Sorts versions so the newest one comes first using version_index and not on id,

src/documents/views.py

@@ -329,9 +329,10 @@ def _get_tantivy_query_and_mode(params):
 def _get_more_like_id(query_params: dict[str, Any], user: User | None) -> int:
     try:
         more_like_doc_id = int(query_params["more_like_id"])
-        more_like_doc = Document.objects.select_related("owner").get(
-            pk=more_like_doc_id,
-        )
+        more_like_doc = Document.objects.select_related(
+            "owner",
+            "root_document__owner",
+        ).get(pk=more_like_doc_id)
     except (TypeError, ValueError, Document.DoesNotExist):
         raise PermissionDenied(_("Invalid more_like_id"))
 
@@ -342,7 +343,8 @@ def _get_more_like_id(query_params: dict[str, Any], user: User | None) -> int:
     ):
         raise PermissionDenied(_("Insufficient permissions."))
 
-    return more_like_doc_id
+    # Only root documents are indexed, a version stands for its root
+    return more_like_doc.root_document_id or more_like_doc.pk
 
 
 class SearchParams(NamedTuple):

src/paperless_ai/indexing.py

@@ -17,6 +17,7 @@ from documents.models import PaperlessTask
 from documents.utils import IterWrapper
 from documents.utils import QuerySetStream
 from documents.utils import identity
+from documents.versioning import root_document_ids
 from paperless.config import AIConfig
 from paperless_ai.db import db_connection_released
 from paperless_ai.embedding import build_llm_index_text
@@ -443,11 +444,11 @@ def update_llm_index(
                 "Skipping LLM index update: migration check deferred; "
                 "will retry next run."
             )
-    documents = Document.objects.select_related(
-        "correspondent",
-        "document_type",
-        "storage_path",
-    ).prefetch_related("tags", "notes", "custom_fields__field")
+    documents = (
+        Document.objects.filter(root_document__isnull=True)
+        .select_related("correspondent", "document_type", "storage_path")
+        .prefetch_related("tags", "notes", "custom_fields__field")
+    )
     no_documents = not documents.exists()
 
     # Fast exit before touching config: nothing to index and no existing index.
@@ -483,7 +484,7 @@ def update_llm_index(
             msg = "LLM index rebuilt successfully."
         else:
             scoped_documents = (
-                documents.filter(id__in=document_ids)
+                documents.filter(id__in=root_document_ids(document_ids))
                 if document_ids is not None
                 else documents
             )
@@ -510,7 +511,12 @@ def update_llm_index(
 
 
 def llm_index_add_or_update_document(document: Document):
-    """Add or atomically replace a document's chunks in the index."""
+    """
+    Add or atomically replace a document's chunks in the index. Only root
+    documents are indexed, so a version is indexed as its root document.
+    """
+    if document.root_document_id is not None:
+        document = document.root_document
     config = AIConfig()
     new_nodes = build_document_node(
         document,

src/paperless_ai/tests/test_ai_indexing.py

@@ -388,6 +388,83 @@ def test_update_llm_index_partial_update(
     assert after[str(doc2.pk)] == before[str(doc2.pk)]
 
 
+@pytest.mark.django_db
+class TestLlmIndexVersions:
+    """The LLM index holds root documents only: a version is indexed as its root."""
+
+    def test_add_or_update_document_indexes_a_version_as_its_root(
+        self,
+        temp_llm_index_dir: Path,
+        mock_embed_model: FakeEmbedding,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - The version is passed to llm_index_add_or_update_document
+        THEN:
+            - Only the root document is in the index
+        """
+        root = DocumentFactory(content="root content")
+        version = DocumentFactory(root_document=root, version_index=1)
+
+        indexing.llm_index_add_or_update_document(version)
+
+        with indexing.get_vector_store() as store:
+            indexed = store.get_modified_times()
+
+        assert set(indexed) == {str(root.pk)}
+
+    def test_rebuild_skips_versions(
+        self,
+        temp_llm_index_dir: Path,
+        mock_embed_model: FakeEmbedding,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - The LLM index is rebuilt
+        THEN:
+            - Only the root document is in the index
+        """
+        root = DocumentFactory()
+        DocumentFactory(root_document=root, version_index=1)
+
+        indexing.update_llm_index(rebuild=True)
+
+        with indexing.get_vector_store() as store:
+            indexed = store.get_modified_times()
+
+        assert set(indexed) == {str(root.pk)}
+
+    def test_incremental_update_by_version_id_refreshes_the_root(
+        self,
+        temp_llm_index_dir: Path,
+        mock_embed_model: FakeEmbedding,
+    ) -> None:
+        """
+        GIVEN:
+            - An indexed root document with a version whose root was modified since
+        WHEN:
+            - An incremental update is scoped to the version's id
+        THEN:
+            - The root's entry is refreshed and no entry exists for the version
+        """
+        root = DocumentFactory()
+        version = DocumentFactory(root_document=root, version_index=1)
+        indexing.update_llm_index(rebuild=True)
+        Document.objects.filter(pk=root.pk).update(modified=timezone.now())
+        root.refresh_from_db()
+
+        indexing.update_llm_index(document_ids=[version.pk])
+
+        with indexing.get_vector_store() as store:
+            indexed = store.get_modified_times()
+
+        assert indexed == {str(root.pk): root.modified.isoformat()}
+
+
 @pytest.mark.django_db
 def test_add_or_update_document_updates_existing_entry(
     temp_llm_index_dir: Path,
@@ -637,6 +714,7 @@ class TestLlmIndexAddOrUpdateDocumentEmptyContent:
 
         doc = MagicMock(spec=Document)
         doc.id = 42
+        doc.root_document_id = None
         # Must not raise
         indexing.llm_index_add_or_update_document(doc)