Not bad catches from Copilot, if a little extra secure

This brings users without an embedding backend configured to closer
parity with those who do.  Reuse the search backend to locate similar
documents and use them to provide the LLM with the better suggestion pool
to draw from

src/paperless_ai/ai_classifier.py

@@ -5,13 +5,14 @@ from django.conf import settings
 from django.contrib.auth.models import User
 
 from documents.models import Document
-from documents.permissions import get_objects_for_user_owner_aware
+from documents.permissions import permitted_object_ids
+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.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
@@ -19,7 +20,9 @@ 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 AssignedMetadata
+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
@@ -39,6 +42,48 @@ 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:
@@ -147,45 +192,54 @@ def get_taxonomy_context(
     user: User | None = None,
     max_docs: int = 5,
 ) -> tuple[TaxonomyCandidates, AssignedMetadata, str]:
-    """One retrieval feeds both taxonomy candidates and RAG text context.
-    On any retrieval failure, degrades to empty candidates/context rather than
-    propagating the exception - a vector-store outage should not block
-    classification, only its RAG-assisted enrichment.
+    """One retrieval feeds both taxonomy candidates and RAG text context. Uses
+    vector similarity when an embedding backend is configured, otherwise
+    falls back to Tantivy full-text "More Like This" similarity - see
+    _fulltext_similar_documents. On any retrieval failure, degrades to empty
+    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.
     """
     assigned = get_assigned_metadata(document, user)
+    ai_config = AIConfig()
     try:
-        # None means "no restriction" to retrieve_similar_nodes. A superuser
-        # (like no user at all) can see every document, so skip materializing
-        # every visible pk into a Python list and passing it through as an IN
-        # filter: for a large library that is a wasted quadratic scan in the
-        # vector store at best, and past ~32,763 documents a hard
-        # sqlite3.OperationalError (SQLite's bound-parameter limit) at worst.
-        # get_objects_for_user_owner_aware() would return every Document for a
-        # superuser anyway (guardian's own with_superuser shortcut), so this
-        # changes nothing about which documents are considered -- only how we
-        # get there.
-        visible_document_ids = (
-            None
-            if user is None or user.is_superuser
-            else list(
-                get_objects_for_user_owner_aware(
-                    user,
-                    "view_document",
-                    Document,
-                ).values_list("pk", flat=True),
+        if ai_config.llm_embedding_backend:
+            # None means "no restriction" to retrieve_similar_nodes. A superuser
+            # (like no user at all) can see every document, so skip materializing
+            # every visible pk into a Python list and passing it through as an IN
+            # filter: for a large library that is a wasted 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_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 None or user.is_superuser
+                else list(permitted_object_ids(user, Document, "view_document"))
+            )
+            nodes = retrieve_similar_nodes(
+                document,
+                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(nodes, user)
+        candidates = build_taxonomy_candidates(similar_documents, user)
 
-        similar_docs = list(
-            Document.objects.filter(pk__in=_node_document_ids(nodes))[:max_docs],
-        )
+        similar_doc_ids = [s["document_id"] for s in similar_documents]
+        docs_by_id = Document.objects.in_bulk(similar_doc_ids)
+        similar_docs = [
+            docs_by_id[doc_id] for doc_id in similar_doc_ids if doc_id in docs_by_id
+        ][:max_docs]
         context_blocks = []
         for similar in similar_docs:
             text = similar.content[:1000] or ""
@@ -193,8 +247,8 @@ def get_taxonomy_context(
             context_blocks.append(f"TITLE: {title}\n{text}")
     except Exception:
         logger.exception(
-            "Failed to retrieve RAG neighbours for document %s; continuing "
-            "without taxonomy candidates or similar-document context.",
+            "Failed to retrieve similar-document context for document %s; "
+            "continuing without taxonomy candidates or similar-document context.",
             document.pk,
         )
         return empty_taxonomy_candidates(), assigned, ""
@@ -277,23 +331,14 @@ def get_ai_document_classification(
 ) -> ClassificationSuggestions:
     ai_config = AIConfig()
 
-    if ai_config.llm_embedding_backend:
-        candidates, assigned, context = get_taxonomy_context(document, user)
-        prompt = build_prompt_with_rag(
-            document,
-            ai_config,
-            candidates=candidates,
-            assigned=assigned,
-            context=context,
-        )
-    else:
-        candidates = empty_taxonomy_candidates()
-        prompt = build_prompt_without_rag(
-            document,
-            ai_config,
-            candidates=candidates,
-            assigned=get_assigned_metadata(document, user),
-        )
+    candidates, assigned, context = get_taxonomy_context(document, user)
+    prompt = build_prompt_with_rag(
+        document,
+        ai_config,
+        candidates=candidates,
+        assigned=assigned,
+        context=context,
+    )
 
     client = AIClient()
     # Hand the pooled DB connection back while the (slow) LLM query runs so it

src/paperless_ai/taxonomy.py

@@ -33,6 +33,11 @@ class TaxonomyCandidate(TypedDict):
     weight: float
 
 
+class SimilarDocument(TypedDict):
+    document_id: int
+    weight: float
+
+
 class TaxonomyCandidates(TypedDict):
     tags: list[TaxonomyCandidate]
     document_types: list[TaxonomyCandidate]
@@ -105,10 +110,10 @@ def get_assigned_metadata(document: Document, user: User | None) -> AssignedMeta
     )
 
 
-def _node_document_weights(nodes: list["NodeWithScore"]) -> dict[int, float]:
-    """document_id -> that node's similarity score, summed if a document_id
-    appears more than once across the retrieved nodes (e.g. multiple chunks
-    of the same source document)."""
+def _node_document_weights(nodes: list["NodeWithScore"]) -> list[SimilarDocument]:
+    """Sum each node's similarity score into its document_id (a document can
+    appear via multiple chunks/nodes) and return one SimilarDocument per
+    distinct document_id."""
     weights: dict[int, float] = defaultdict(float)
     for node in nodes:
         document_id = node.metadata.get("document_id")
@@ -121,7 +126,10 @@ def _node_document_weights(nodes: list["NodeWithScore"]) -> dict[int, float]:
             weights[int(document_id)] += float(node.score or 0.0)
         except (TypeError, ValueError):  # pragma: no cover
             continue
-    return weights
+    return [
+        SimilarDocument(document_id=document_id, weight=weight)
+        for document_id, weight in weights.items()
+    ]
 
 
 def _visible_ranked_candidates(
@@ -157,21 +165,26 @@ def _visible_ranked_candidates(
 
 
 def build_taxonomy_candidates(
-    nodes: list["NodeWithScore"],
+    similar_documents: list[SimilarDocument],
     user: User | None,
 ) -> TaxonomyCandidates:
-    """Resolve each neighbour node's document_id to a live Document, read its
-    *current* tags/type/correspondent/storage_path via the ORM (never the
-    possibly-stale names cached in vector-index node metadata), weight each
-    distinct taxonomy object by aggregate neighbour similarity, permission-filter
+    """Resolve each similar document's id to a live Document, read its
+    *current* tags/type/correspondent/storage_path via the ORM (never any
+    possibly-stale names an adapter's source might have cached), weight each
+    distinct taxonomy object by aggregate similarity weight, permission-filter
     against what ``user`` can see, and return each category ranked by weight
-    and capped.
+    and capped. ``similar_documents`` may come from either the vector-RAG
+    adapter or the full-text fallback adapter - both produce this same shape.
     """
-
-    document_weights = _node_document_weights(nodes)
-    if not document_weights:
+    if not similar_documents:
         return empty_taxonomy_candidates()
 
+    # Both adapters guarantee at most one SimilarDocument per document_id, so
+    # this never silently drops a duplicate's weight.
+    document_weights: dict[int, float] = {
+        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
     # below via their *_id columns (neighbour.document_type_id, etc.), which

src/paperless_ai/tests/test_ai_classifier.py

@@ -1,3 +1,4 @@
+from collections.abc import Generator
 from types import SimpleNamespace
 from unittest.mock import MagicMock
 from unittest.mock import patch
@@ -7,10 +8,13 @@ import pytest_mock
 from django.test import override_settings
 
 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 _restrict_to_shown_candidates
 from paperless_ai.ai_classifier import build_localization_prompt
 from paperless_ai.ai_classifier import build_prompt_with_rag
@@ -20,6 +24,7 @@ from paperless_ai.ai_classifier import get_language_name
 from paperless_ai.ai_classifier import get_taxonomy_context
 from paperless_ai.base_model import ClassificationSuggestions
 from paperless_ai.base_model import TaxonomyChoiceDict
+from paperless_ai.taxonomy import SimilarDocument
 from paperless_ai.taxonomy import TaxonomyCandidate
 from paperless_ai.taxonomy import TaxonomyCandidates
 from paperless_ai.taxonomy import empty_taxonomy_candidates
@@ -204,12 +209,10 @@ 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_without_rag")
-@patch("paperless_ai.ai_classifier.AIConfig")
+@patch("paperless_ai.ai_classifier.build_prompt_with_rag")
 @override_settings(LLM_BACKEND="ollama", LLM_MODEL="some_model")
-def test_use_without_rag_if_not_configured(
-    mock_ai_config,
-    mock_build_prompt_without_rag,
+def test_use_rag_prompt_even_without_embedding_backend(
+    mock_build_prompt_with_rag,
     mock_run_llm_query,
     mock_document,
 ):
@@ -219,13 +222,13 @@ def test_use_without_rag_if_not_configured(
     WHEN:
         - get_ai_document_classification() is called
     THEN:
-        - The non-RAG prompt builder is used
+        - The RAG-context prompt builder is still used (fed by the full-text
+          fallback's context/candidates instead of the vector store's)
     """
-    mock_ai_config.return_value.llm_embedding_backend = None
-    mock_build_prompt_without_rag.return_value = "Prompt without RAG"
+    mock_build_prompt_with_rag.return_value = "Prompt with RAG"
     mock_run_llm_query.return_value = NESTED_SUGGESTIONS
     get_ai_document_classification(mock_document)
-    mock_build_prompt_without_rag.assert_called_once()
+    mock_build_prompt_with_rag.assert_called_once()
 
 
 @pytest.mark.django_db
@@ -303,6 +306,7 @@ 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:
@@ -344,6 +348,7 @@ 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_no_similar_docs():
     """
     GIVEN:
@@ -367,6 +372,67 @@ 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
@@ -379,6 +445,7 @@ class TestGetTaxonomyContextVisibility:
     """
 
     @pytest.mark.django_db
+    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
     def test_skips_permission_lookup_for_superuser(
         self,
         mocker: pytest_mock.MockerFixture,
@@ -397,17 +464,18 @@ class TestGetTaxonomyContextVisibility:
             "paperless_ai.ai_classifier.retrieve_similar_nodes",
             return_value=[],
         )
-        mock_get_objects = mocker.patch(
-            "paperless_ai.ai_classifier.get_objects_for_user_owner_aware",
+        mock_permitted = mocker.patch(
+            "paperless_ai.ai_classifier.permitted_object_ids",
         )
         user = UserFactory.create(is_superuser=True)
 
         get_taxonomy_context(document, user)
 
-        mock_get_objects.assert_not_called()
+        mock_permitted.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,
@@ -426,16 +494,17 @@ class TestGetTaxonomyContextVisibility:
             "paperless_ai.ai_classifier.retrieve_similar_nodes",
             return_value=[],
         )
-        mock_get_objects = mocker.patch(
-            "paperless_ai.ai_classifier.get_objects_for_user_owner_aware",
+        mock_permitted = mocker.patch(
+            "paperless_ai.ai_classifier.permitted_object_ids",
         )
 
         get_taxonomy_context(document, None)
 
-        mock_get_objects.assert_not_called()
+        mock_permitted.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,
@@ -446,7 +515,7 @@ class TestGetTaxonomyContextVisibility:
         WHEN:
             - get_taxonomy_context() is called
         THEN:
-            - The user's visible document ids are looked up and passed to
+            - The user's permitted document ids are looked up and passed to
               retrieve_similar_nodes() as a restriction
         """
         document = DocumentFactory.create(content="Some content")
@@ -454,21 +523,186 @@ class TestGetTaxonomyContextVisibility:
             "paperless_ai.ai_classifier.retrieve_similar_nodes",
             return_value=[],
         )
-        mock_queryset = mocker.MagicMock()
-        mock_queryset.values_list.return_value = [1, 2, 3]
-        mock_get_objects = mocker.patch(
-            "paperless_ai.ai_classifier.get_objects_for_user_owner_aware",
-            return_value=mock_queryset,
+        mock_permitted = mocker.patch(
+            "paperless_ai.ai_classifier.permitted_object_ids",
+            return_value=[1, 2, 3],
         )
         user = UserFactory.create(is_superuser=False)
 
         get_taxonomy_context(document, user)
 
-        mock_get_objects.assert_called_once_with(user, "view_document", Document)
+        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
+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]
+
+
+@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):
     """
@@ -495,6 +729,7 @@ 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_taxonomy.py

@@ -1,5 +1,4 @@
 import json
-from types import SimpleNamespace
 
 import pytest
 import pytest_mock
@@ -11,6 +10,7 @@ from documents.tests.factories import StoragePathFactory
 from documents.tests.factories import TagFactory
 from documents.tests.factories import UserFactory
 from paperless_ai.taxonomy import AssignedMetadata
+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
@@ -132,9 +132,8 @@ class TestGetAssignedMetadata:
         assert result["tags"] == ["Owned By Someone Else"]
 
 
-def make_node(document_id: int, score: float) -> SimpleNamespace:
-    """A stand-in for NodeWithScore: only ``.metadata``/``.score`` are read."""
-    return SimpleNamespace(metadata={"document_id": str(document_id)}, score=score)
+def make_similar(document_id: int, weight: float) -> SimilarDocument:
+    return SimilarDocument(document_id=document_id, weight=weight)
 
 
 @pytest.mark.django_db
@@ -170,9 +169,9 @@ class TestBuildTaxonomyCandidates:
         doc_a.tags.add(tag)
         doc_b = DocumentFactory.create()
         doc_b.tags.add(tag)
-        nodes = [make_node(doc_a.pk, 0.9), make_node(doc_b.pk, 0.4)]
+        similar_documents = [make_similar(doc_a.pk, 0.9), make_similar(doc_b.pk, 0.4)]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["tags"]) == 1
         assert result["tags"][0]["id"] == tag.pk
@@ -197,9 +196,9 @@ class TestBuildTaxonomyCandidates:
         document.tags.add(tag)
         tag.name = "New Name"
         tag.save()
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert result["tags"][0]["name"] == "New Name"
 
@@ -219,9 +218,9 @@ class TestBuildTaxonomyCandidates:
         document = DocumentFactory.create()
         document.tags.add(tag)
         tag.delete()
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert result["tags"] == []
 
@@ -240,9 +239,12 @@ class TestBuildTaxonomyCandidates:
         strong_doc.tags.add(strong_tag)
         weak_doc = DocumentFactory.create()
         weak_doc.tags.add(weak_tag)
-        nodes = [make_node(strong_doc.pk, 0.9), make_node(weak_doc.pk, 0.1)]
+        similar_documents = [
+            make_similar(strong_doc.pk, 0.9),
+            make_similar(weak_doc.pk, 0.1),
+        ]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert [c["name"] for c in result["tags"]] == ["Strong", "Weak"]
 
@@ -258,9 +260,9 @@ class TestBuildTaxonomyCandidates:
         document = DocumentFactory.create()
         for i in range(15):
             document.tags.add(TagFactory.create(name=f"Tag{i}"))
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["tags"]) == 10
 
@@ -274,12 +276,12 @@ class TestBuildTaxonomyCandidates:
             - Only 5 correspondents are returned
         """
         correspondents = CorrespondentFactory.create_batch(7)
-        nodes = [
-            make_node(DocumentFactory.create(correspondent=c).pk, 0.5)
+        similar_documents = [
+            make_similar(DocumentFactory.create(correspondent=c).pk, 0.5)
             for c in correspondents
         ]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["correspondents"]) == 5
 
@@ -294,9 +296,9 @@ class TestBuildTaxonomyCandidates:
         """
         document_type = DocumentTypeFactory.create(name="Invoice")
         document = DocumentFactory.create(document_type=document_type)
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["document_types"]) == 1
         assert result["document_types"][0]["id"] == document_type.pk
@@ -312,12 +314,12 @@ class TestBuildTaxonomyCandidates:
             - Only 5 document_types are returned
         """
         document_types = DocumentTypeFactory.create_batch(7)
-        nodes = [
-            make_node(DocumentFactory.create(document_type=dt).pk, 0.5)
+        similar_documents = [
+            make_similar(DocumentFactory.create(document_type=dt).pk, 0.5)
             for dt in document_types
         ]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["document_types"]) == 5
 
@@ -332,9 +334,9 @@ class TestBuildTaxonomyCandidates:
         """
         storage_path = StoragePathFactory.create(name="Invoices")
         document = DocumentFactory.create(storage_path=storage_path)
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(document.pk, 0.5)]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["storage_paths"]) == 1
         assert result["storage_paths"][0]["id"] == storage_path.pk
@@ -350,12 +352,12 @@ class TestBuildTaxonomyCandidates:
             - Only 5 storage_paths are returned
         """
         storage_paths = StoragePathFactory.create_batch(7)
-        nodes = [
-            make_node(DocumentFactory.create(storage_path=sp).pk, 0.5)
+        similar_documents = [
+            make_similar(DocumentFactory.create(storage_path=sp).pk, 0.5)
             for sp in storage_paths
         ]
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert len(result["storage_paths"]) == 5
 
@@ -375,14 +377,14 @@ class TestBuildTaxonomyCandidates:
         tag = TagFactory.create(name="Restricted")
         document = DocumentFactory.create()
         document.tags.add(tag)
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(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(nodes, user=user)
+        result = build_taxonomy_candidates(similar_documents, user=user)
 
         assert result["tags"] == []
 
@@ -412,10 +414,10 @@ class TestBuildTaxonomyCandidates:
         tag.save()
         document = DocumentFactory.create()
         document.tags.add(tag)
-        nodes = [make_node(document.pk, 0.5)]
+        similar_documents = [make_similar(document.pk, 0.5)]
         spy = mocker.patch("documents.permissions.permitted_object_ids")
 
-        result = build_taxonomy_candidates(nodes, user=None)
+        result = build_taxonomy_candidates(similar_documents, user=None)
 
         assert result["tags"][0]["name"] == "Owned"
         spy.assert_not_called()