Fix: use a root document's effective content for the LLM index and prompts

The search index and the document views use the newest version's content as a
root document's content, but everything the LLM side reads used the root's own
content: the text embedded in the LLM index, the content in the classification
prompt, the context blocks from similar documents, and the text that is
searched for similar documents. After a document was replaced by a new
version the LLM kept answering from the old text.

Read get_effective_content() in those four places. The index update annotates
the content in the query like the search index does, and the similar-document
lookup annotates it for the documents it fetches, so no query is made per
document. Entries already in the LLM index keep the old text until the
document next changes or the index is rebuilt.

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(
+                documents = (
-                    "correspondent",
+                    Document.objects.filter(root_document__isnull=True)
-                    "document_type",
+                    .select_related(
-                    "storage_path",
+                        "correspondent",
-                    "owner",
+                        "document_type",
-                ).prefetch_related(
+                        "storage_path",
-                    "tags",
+                        "owner",
-                    "notes__user",
+                    )
-                    "custom_fields__field",
+                    .prefetch_related(
-                    "versions",
+                        "tags",
-                    "barcodes",
+                        "notes__user",
-                    "versions__barcodes",
+                        "custom_fields__field",
+                        "versions",
+                        "barcodes",
+                        "versions__barcodes",
+                    )
                 )
                 total = documents.count()
                 rebuild_kwargs = {}

src/documents/permissions.py

@@ -8,9 +8,11 @@ from django.contrib.contenttypes.models import ContentType
 from django.db.models import Case
 from django.db.models import CharField
 from django.db.models import Count
+from django.db.models import Exists
 from django.db.models import F
 from django.db.models import IntegerField
 from django.db.models import Model
+from django.db.models import OuterRef
 from django.db.models import Q
 from django.db.models import QuerySet
 from django.db.models import Value
@@ -367,18 +369,14 @@ def permitted_object_ids(
     visible exactly when its parent is, judged by the parent's owner and
     grants, so the row's own owner and grants are ignored.
 
-    Guardian stores ``object_pk`` as a string, so the row key is cast to a
+    Guardian stores ``object_pk`` as a string, so each grant is an ``EXISTS``
-    string and tested against the user's and groups' grants with a single
+    keyed on the row's id cast to a string, which can use guardian's
-    uncorrelated ``IN``. Postgres and SQLite build that set once. MariaDB
+    (user, permission, object_pk) unique index. Casting every ``object_pk`` to
-    evaluates it as an index probe per row, which is cheap because the
+    an integer for an ``id IN (...)`` is not indexable, and MariaDB cannot
-    lookups use guardian's unique indexes. Casting every ``object_pk`` to an
+    materialize it inside the owner ``OR``, so it re-scans the user's grants
-    integer instead cannot use an index, and MariaDB cannot materialize it
+    for every row. The user's groups are matched with an ``IN`` subquery
-    inside the owner ``OR``, so it re-scans the user's grants for every row.
+    rather than a join through the membership table, which SQLite plans badly
-    A correlated ``EXISTS`` per grant fixes MariaDB too, but Postgres and
+    once the grant tables grow.
-    SQLite re-run it for every row and end up slower than the original. The
-    user's groups are matched with an ``IN`` subquery rather than a join
-    through the membership table, which SQLite plans badly once the grant
-    tables grow.
     """
     has_soft_delete = hasattr(model, "global_objects")
     manager = (
@@ -424,26 +422,25 @@ def permitted_object_ids(
         "permission__content_type": content_type,
     }
 
-    # Both grant sets are compared to the row key as strings, exactly as
+    key_as_text = Cast(OuterRef(key_field), CharField(max_length=64))
-    # guardian stores them, and are uncorrelated, so each engine can build the
+    granted_to_user = Exists(
-    # set once instead of probing per row.
+        UserObjectPermission.objects.filter(
-    user_keys = UserObjectPermission.objects.filter(
+            user=user,
-        user=user,
+            object_pk=key_as_text,
-        **perm_filter,
+            **perm_filter,
-    ).values_list("object_pk", flat=True)
+        ),
-    group_keys = GroupObjectPermission.objects.filter(
-        group_id__in=user.groups.values("id"),
-        **perm_filter,
-    ).values_list("object_pk", flat=True)
-    permitted_keys = user_keys.union(group_keys, all=True)
-
-    return (
-        base_qs.annotate(permitted_key=Cast(key_field, CharField(max_length=64)))
-        .filter(
-            Q(**{owner_field: user.pk}) | unowned | Q(permitted_key__in=permitted_keys),
-        )
-        .values_list("id", flat=True)
     )
+    granted_to_group = Exists(
+        GroupObjectPermission.objects.filter(
+            group_id__in=user.groups.values("id"),
+            object_pk=key_as_text,
+            **perm_filter,
+        ),
+    )
+
+    return base_qs.filter(
+        Q(**{owner_field: user.pk}) | unowned | granted_to_user | granted_to_group,
+    ).values_list("id", flat=True)
 
 
 ModelT = TypeVar("ModelT", bound=Model)

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",

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:
+            if document.root_document_id is not None:
-                clear_document_caches(root_document.pk)
+                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."""
 

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 version's content is updated, not the root's
-            - The root document is indexed rather than the version
+            - 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]
+        mock_get_backend.return_value.add_or_update.assert_called_once()
-        self.assertEqual(indexed.pk, root.pk)
         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(
+        more_like_doc = Document.objects.select_related(
-            pk=more_like_doc_id,
+            "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/ai_classifier.py

@@ -4,8 +4,10 @@ from django.conf import settings
 from django.contrib.auth.models import User
 
 from documents.models import Document
-from documents.permissions import permitted_document_ids
+from documents.permissions import permitted_object_ids
+from documents.permissions import restrict_queryset_to_visible
 from documents.permissions import user_is_unrestricted
+from documents.versioning import annotate_effective_content
 from paperless.config import AIConfig
 from paperless_ai.base_model import ClassificationSuggestions
 from paperless_ai.base_model import TaxonomyChoiceDict
@@ -53,8 +55,8 @@ def _fulltext_similar_documents(
     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
-    permitted_document_ids() since Tantivy's indexed permission fields lag the
+    restrict_queryset_to_visible() since Tantivy's indexed permission fields
-    DB via async reindexing and judge a version by its own owner.
+    lag the DB via async reindexing.
     """
     from documents.search import get_backend
 
@@ -68,9 +70,10 @@ def _fulltext_similar_documents(
     )
     if not unrestricted:
         allowed_ids = set(
-            Document.objects.filter(
+            restrict_queryset_to_visible(
-                pk__in=similar_ids,
+                Document.objects.filter(pk__in=similar_ids),
-                id__in=permitted_document_ids(user),
+                user,
+                "view_document",
             ).values_list("pk", flat=True),
         )
         similar_ids = [doc_id for doc_id in similar_ids if doc_id in allowed_ids]
@@ -111,7 +114,7 @@ def build_prompt_without_rag(
 ) -> str:
     filename = document.filename or ""
     content = truncate_content(
-        document.content[:4000] or "",
+        (document.get_effective_content() or "")[:4000],
         chunk_size=config.llm_embedding_chunk_size,
         context_size=config.llm_context_size,
     )
@@ -198,13 +201,13 @@ def get_taxonomy_context(
             # quadratic scan in the vector store at best, and past ~32,763
             # documents a hard sqlite3.OperationalError (SQLite's
             # bound-parameter limit) at worst.
-            # permitted_document_ids() has its own superuser shortcut that would
+            # permitted_object_ids() has its own superuser shortcut that would
             # return every Document's id anyway, so this changes nothing about
             # which documents are considered -- only how we get there.
             visible_document_ids = (
                 None
                 if user_is_unrestricted(user)
-                else list(permitted_document_ids(user))
+                else list(permitted_object_ids(user, Document, "view_document"))
             )
             nodes = retrieve_similar_nodes(
                 document,
@@ -225,7 +228,9 @@ def get_taxonomy_context(
 
         # similar_documents is already ordered by descending weight; don't lose it.
         similar_document_ids = [s["document_id"] for s in similar_documents]
-        similar_documents_by_id = Document.objects.in_bulk(similar_document_ids)
+        similar_documents_by_id = annotate_effective_content(
+            Document.objects.all(),
+        ).in_bulk(similar_document_ids)
         similar_docs = [
             similar_documents_by_id[document_id]
             for document_id in similar_document_ids
@@ -233,7 +238,7 @@ def get_taxonomy_context(
         ][:max_docs]
         context_blocks = []
         for similar in similar_docs:
-            text = similar.content[:1000] or ""
+            text = (similar.get_effective_content() or "")[:1000]
             title = similar.title or similar.filename or "Untitled"
             context_blocks.append(f"TITLE: {title}\n{text}")
     except Exception:

src/paperless_ai/embedding.py

@@ -135,6 +135,6 @@ def build_llm_index_text(doc: Document) -> str:
         lines.append(f"Custom Field - {instance.field.name}: {instance}")
 
     lines.append("\nContent:\n")
-    lines.append(doc.content or "")
+    lines.append(doc.get_effective_content() or "")
 
     return _normalize_llm_index_text("\n".join(lines))

src/paperless_ai/indexing.py

@@ -17,6 +17,8 @@ from documents.models import PaperlessTask
 from documents.utils import IterWrapper
 from documents.utils import QuerySetStream
 from documents.utils import identity
+from documents.versioning import annotate_effective_content
+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 +445,11 @@ def update_llm_index(
                 "Skipping LLM index update: migration check deferred; "
                 "will retry next run."
             )
-    documents = Document.objects.select_related(
+    documents = annotate_effective_content(
-        "correspondent",
+        Document.objects.filter(root_document__isnull=True)
-        "document_type",
+        .select_related("correspondent", "document_type", "storage_path")
-        "storage_path",
+        .prefetch_related("tags", "notes", "custom_fields__field"),
-    ).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 +485,7 @@ 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 +512,12 @@ 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,
@@ -688,7 +695,7 @@ def retrieve_similar_nodes(
     )
 
     query_text = truncate_embedding_query(
-        (document.title or "") + "\n" + (document.content or ""),
+        (document.title or "") + "\n" + (document.get_effective_content() or ""),
         chunk_size=config.llm_embedding_chunk_size,
     )
     # Hold the shared read lock for the whole retrieval so the connection is

src/paperless_ai/tests/test_ai_classifier.py

@@ -55,6 +55,7 @@ def mock_document():
     doc.storage_path = None
     doc.archive_serial_number = "12345"
     doc.content = "This is the document content."
+    doc.get_effective_content.return_value = "This is the document content."
 
     cf1 = MagicMock(__str__=lambda x: "Value1")
     cf1.field = MagicMock()
@@ -434,6 +435,55 @@ def test_get_taxonomy_context_preserves_similarity_order_and_distinct_documents(
     )
 
 
+@pytest.mark.django_db
+class TestClassifierEffectiveContent:
+    """A root document's text for the LLM is its newest version's content."""
+
+    @staticmethod
+    def _root_with_version() -> Document:
+        root = DocumentFactory(title="Statement", content="stale text")
+        DocumentFactory(root_document=root, version_index=1, content="latest text")
+        return root
+
+    def test_prompt_uses_the_newest_versions_content(self) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - The classification prompt is built for the root
+        THEN:
+            - It contains the newest version's content
+        """
+        prompt = build_prompt_without_rag(self._root_with_version(), AIConfig())
+
+        assert "latest text" in prompt
+        assert "stale text" not in prompt
+
+    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
+    def test_similar_document_context_uses_the_newest_versions_content(self) -> None:
+        """
+        GIVEN:
+            - A similar root document with a version
+        WHEN:
+            - The similar-document context is built
+        THEN:
+            - It contains the newest version's content
+        """
+        similar = self._root_with_version()
+        document = DocumentFactory(content="Some content")
+        fake_nodes = [
+            SimpleNamespace(metadata={"document_id": str(similar.pk)}, score=0.9),
+        ]
+
+        with patch(
+            "paperless_ai.ai_classifier.retrieve_similar_nodes",
+            return_value=fake_nodes,
+        ):
+            _candidates, context = get_taxonomy_context(document, user=None)
+
+        assert context == "TITLE: Statement\nlatest text"
+
+
 @pytest.mark.django_db
 @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
 def test_get_taxonomy_context_no_similar_docs():
@@ -552,7 +602,7 @@ class TestGetTaxonomyContextVisibility:
             return_value=[],
         )
         mock_permitted = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_document_ids",
+            "paperless_ai.ai_classifier.permitted_object_ids",
         )
         user = UserFactory.create(is_superuser=True)
 
@@ -582,7 +632,7 @@ class TestGetTaxonomyContextVisibility:
             return_value=[],
         )
         mock_permitted = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_document_ids",
+            "paperless_ai.ai_classifier.permitted_object_ids",
         )
 
         get_taxonomy_context(document, None)
@@ -611,55 +661,16 @@ class TestGetTaxonomyContextVisibility:
             return_value=[],
         )
         mock_permitted = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_document_ids",
+            "paperless_ai.ai_classifier.permitted_object_ids",
             return_value=[1, 2, 3],
         )
         user = UserFactory.create(is_superuser=False)
 
         get_taxonomy_context(document, user)
 
-        mock_permitted.assert_called_once_with(user)
+        mock_permitted.assert_called_once_with(user, Document, "view_document")
         assert mock_retrieve.call_args.kwargs["document_ids"] == [1, 2, 3]
 
-    @pytest.mark.django_db
-    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
-    def test_version_of_private_root_is_not_visible(
-        self,
-        mocker: pytest_mock.MockerFixture,
-    ) -> None:
-        """
-        GIVEN:
-            - A private root document owned by someone else
-            - A version of it whose own owner is unset, as when the root
-              changed hands after the version was created
-        WHEN:
-            - get_taxonomy_context() is called for a non-superuser
-        THEN:
-            - Neither the root nor the version is in the visible ids passed to
-              retrieve_similar_nodes(), since a version follows its root
-        """
-        owner = UserFactory.create()
-        viewer = UserFactory.create(is_superuser=False)
-        root = DocumentFactory.create(content="private", owner=owner)
-        version = DocumentFactory.create(
-            content="private",
-            owner=None,
-            root_document=root,
-            version_index=1,
-        )
-        source = DocumentFactory.create(content="Some content", owner=viewer)
-        mock_retrieve = mocker.patch(
-            "paperless_ai.ai_classifier.retrieve_similar_nodes",
-            return_value=[],
-        )
-
-        get_taxonomy_context(source, viewer)
-
-        visible = mock_retrieve.call_args.kwargs["document_ids"]
-        assert source.pk in visible
-        assert root.pk not in visible
-        assert version.pk not in visible
-
 
 @pytest.mark.django_db
 class TestFulltextSimilarDocuments:
@@ -842,7 +853,7 @@ class TestFulltextSimilarDocuments:
             - _fulltext_similar_documents() is called with that user
         THEN:
             - Only the still-permitted document is returned - the DB
-              re-check via permitted_document_ids() must catch the
+              re-check via restrict_queryset_to_visible() must catch the
               document Tantivy's stale index still thinks is visible
         """
         owner = UserFactory.create()
@@ -873,45 +884,6 @@ class TestFulltextSimilarDocuments:
 
         assert [s["document_id"] for s in result] == [permitted.pk]
 
-    def test_excludes_version_of_private_root_for_regular_user(
-        self,
-        fulltext_backend: TantivyBackend,
-    ) -> None:
-        """
-        GIVEN:
-            - A regular user and a private root owned by someone else
-            - A version of that root with no owner of its own, which the
-              Tantivy index therefore treats as visible to everyone
-        WHEN:
-            - _fulltext_similar_documents() is called with that user
-        THEN:
-            - The version is not returned, since the DB re-check judges it by
-              its root
-        """
-        owner = UserFactory.create()
-        viewer = UserFactory.create(is_superuser=False)
-        source = DocumentFactory.create(
-            content="shared content phrase",
-            owner=viewer,
-        )
-        root = DocumentFactory.create(
-            content="shared content phrase",
-            owner=owner,
-        )
-        version = DocumentFactory.create(
-            content="shared content phrase",
-            owner=None,
-            root_document=root,
-            version_index=1,
-        )
-        fulltext_backend.add_or_update(source)
-        fulltext_backend.add_or_update(root)
-        fulltext_backend.add_or_update(version)
-
-        result = _fulltext_similar_documents(source, user=viewer, top_k=5)
-
-        assert version.pk not in [s["document_id"] for s in result]
-
 
 @pytest.mark.django_db
 @override_settings(LLM_EMBEDDING_BACKEND="huggingface")

src/paperless_ai/tests/test_ai_indexing.py

@@ -388,6 +388,108 @@ 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_rebuild_indexes_the_newest_versions_content(
+        self,
+        temp_llm_index_dir: Path,
+        mock_embed_model: FakeEmbedding,
+        mocker: pytest_mock.MockerFixture,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document with a version
+        WHEN:
+            - The LLM index is rebuilt
+        THEN:
+            - The root's text for the index is the version's content, answered
+              from the query rather than a query per document
+        """
+        root = DocumentFactory(content="stale text")
+        DocumentFactory(root_document=root, version_index=1, content="latest text")
+        spy = mocker.spy(indexing, "build_document_node")
+
+        indexing.update_llm_index(rebuild=True)
+
+        indexed = spy.call_args.args[0]
+        assert indexed.pk == root.pk
+        assert indexed.effective_content == "latest text"
+
+    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 +739,7 @@ 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)
 

src/paperless_ai/tests/test_embedding.py

@@ -12,6 +12,7 @@ from paperless_ai.embedding import _normalize_llm_index_text
 from paperless_ai.embedding import build_llm_index_text
 from paperless_ai.embedding import get_configured_model_name
 from paperless_ai.embedding import get_embedding_model
+from paperless_testing.factories import DocumentFactory
 
 
 @pytest.fixture
@@ -46,6 +47,7 @@ def mock_document():
     doc.correspondent.name = "Test Correspondent"
     doc.archive_serial_number = "12345"
     doc.content = "This is the document content."
+    doc.get_effective_content.return_value = "This is the document content."
 
     cf1 = MagicMock(__str__=lambda x: "Value1")
     cf1.field = MagicMock()
@@ -280,7 +282,7 @@ def test_build_llm_index_text(mock_document):
 
 
 def test_build_llm_index_text_normalizes_ocr_punctuation_runs(mock_document):
-    mock_document.content = (
+    mock_document.get_effective_content.return_value = (
         "Introduction ................................................ 7\n"
         "Hardware Limitation ________________________________________ 9\n"
         "Keep short punctuation like INV-100 and ellipses..."
@@ -294,6 +296,43 @@ def test_build_llm_index_text_normalizes_ocr_punctuation_runs(mock_document):
     assert "ellipses..." in result
 
 
+@pytest.mark.django_db
+class TestBuildLlmIndexTextVersions:
+    """A root document is indexed with its effective content, like in the search index."""
+
+    def test_root_uses_the_newest_versions_content(self) -> None:
+        """
+        GIVEN:
+            - A root document with two versions
+        WHEN:
+            - The LLM index text is built for the root
+        THEN:
+            - It contains the newest version's content and not the others'
+        """
+        root = DocumentFactory(content="stale text")
+        DocumentFactory(root_document=root, version_index=1, content="older text")
+        DocumentFactory(root_document=root, version_index=2, content="latest text")
+
+        text = build_llm_index_text(root)
+
+        assert "latest text" in text
+        assert "stale text" not in text
+        assert "older text" not in text
+
+    def test_root_without_versions_uses_its_own_content(self) -> None:
+        """
+        GIVEN:
+            - A root document without versions
+        WHEN:
+            - The LLM index text is built for it
+        THEN:
+            - It contains the document's own content
+        """
+        root = DocumentFactory(content="own text")
+
+        assert "own text" in build_llm_index_text(root)
+
+
 def test_normalize_llm_index_text_collapses_ocr_leaders_without_joining_lines():
     assert _normalize_llm_index_text("A........B\nC____D----E") == "A B\nC D E"