Fix: 3.1.0 llm suggestion raw cache user scoping (#13849)

src/documents/caching.py

@@ -1,7 +1,9 @@
 from __future__ import annotations
 
+import hashlib
 import logging
 import pickle
+import uuid
 from binascii import hexlify
 from collections import OrderedDict
 from dataclasses import dataclass
@@ -55,6 +57,8 @@ LLM_CACHE_CLASSIFIER_VERSION: Final[int] = 1001
 CACHE_1_MINUTE: Final[int] = 60
 CACHE_5_MINUTES: Final[int] = 5 * CACHE_1_MINUTE
 CACHE_50_MINUTES: Final[int] = 50 * CACHE_1_MINUTE
+# Deliberately longer than any entry it names
+LLM_CACHE_GENERATION_TIMEOUT: Final[int] = 2 * CACHE_50_MINUTES
 
 read_cache = caches["read-cache"]
 
@@ -206,12 +210,40 @@ def refresh_suggestions_cache(
     cache.touch(doc_key, timeout)
 
 
+def invalidate_suggestions_cache(document_id: int) -> None:
+    """Invalidate classifier-generated suggestions for a document."""
+    cache.delete(get_suggestion_cache_key(document_id))
+
+
+def _llm_generation_key(document_id: int) -> str:
+    return f"{get_suggestion_cache_key(document_id)}_llm_generation"
+
+
+def _llm_variant_key(document_id: int, backend: str) -> str:
+    """Cache key for one LLM configuration and permission scope.
+
+    ``backend`` identifies the variant - model, endpoint, output language and
+    requesting user.
+
+    Generating the token on first use lets invalidate_llm_suggestions_cache()
+    be no-op for documents that never had AI suggestions.
+    """
+    generation_key = _llm_generation_key(document_id)
+    generation = cache.get_or_set(
+        generation_key,
+        lambda: uuid.uuid4().hex,
+        timeout=LLM_CACHE_GENERATION_TIMEOUT,
+    )
+    cache.touch(generation_key, LLM_CACHE_GENERATION_TIMEOUT)
+    backend_hash = hashlib.sha256(backend.encode()).hexdigest()[:16]
+    return f"{get_suggestion_cache_key(document_id)}_llm_{generation}_{backend_hash}"
+
+
 def get_llm_suggestion_cache(
     document_id: int,
     backend: str,
 ) -> SuggestionCacheData | None:
-    doc_key = get_suggestion_cache_key(document_id)
-    data: SuggestionCacheData = cache.get(doc_key)
+    data: SuggestionCacheData = cache.get(_llm_variant_key(document_id, backend))
 
     if (
         data
@@ -234,9 +266,8 @@ def set_llm_suggestions_cache(
     Cache LLM-generated suggestions using a backend-specific identifier
     (e.g. 'openai-like:gpt-4').
     """
-    doc_key = get_suggestion_cache_key(document_id)
     cache.set(
-        doc_key,
+        _llm_variant_key(document_id, backend),
         SuggestionCacheData(
             classifier_version=LLM_CACHE_CLASSIFIER_VERSION,
             classifier_hash=backend,
@@ -246,17 +277,31 @@ def set_llm_suggestions_cache(
     )
 
 
+def refresh_llm_suggestions_cache(
+    document_id: int,
+    backend: str,
+    *,
+    timeout: int = CACHE_50_MINUTES,
+) -> None:
+    """
+    Refreshes the expiration of one cached LLM suggestion variant.
+    """
+    cache.touch(_llm_variant_key(document_id, backend), timeout)
+
+
 def invalidate_llm_suggestions_cache(
     document_id: int,
 ) -> None:
     """
-    Invalidate the LLM suggestions cache for a specific document and backend.
+    Invalidate every LLM suggestion variant for a document.
     """
-    doc_key = get_suggestion_cache_key(document_id)
-    data: SuggestionCacheData = cache.get(doc_key)
-
-    if data:
-        cache.delete(doc_key)
+    generation_key = _llm_generation_key(document_id)
+    if cache.get(generation_key) is not None:
+        cache.set(
+            generation_key,
+            uuid.uuid4().hex,
+            timeout=LLM_CACHE_GENERATION_TIMEOUT,
+        )
 
 
 def get_metadata_cache_key(document_id: int) -> str:
@@ -357,3 +402,4 @@ def clear_document_caches(document_id: int) -> None:
             get_thumbnail_modified_key(document_id),
         ],
     )
+    invalidate_llm_suggestions_cache(document_id)

src/documents/signals/handlers.py

@@ -32,6 +32,7 @@ from rest_framework import serializers
 from documents import matching
 from documents.caching import clear_document_caches
 from documents.caching import invalidate_llm_suggestions_cache
+from documents.caching import invalidate_suggestions_cache
 from documents.data_models import ConsumableDocument
 from documents.file_handling import create_source_path_directory
 from documents.file_handling import delete_empty_directories
@@ -740,9 +741,9 @@ def cleanup_custom_field_deletion(sender, instance: CustomField, **kwargs) -> No
 @receiver(models.signals.post_save, sender=Document)
 def update_llm_suggestions_cache(sender, instance, **kwargs):
     """
-    Invalidate the LLM suggestions cache when a document is saved.
+    Invalidate suggestions caches when a document is saved.
     """
-    # Invalidate the cache for the document
+    invalidate_suggestions_cache(instance.pk)
     invalidate_llm_suggestions_cache(instance.pk)
 
 

src/documents/tests/test_views.py

@@ -9,6 +9,7 @@ from django.conf import settings
 from django.contrib.auth.models import Group
 from django.contrib.auth.models import Permission
 from django.contrib.auth.models import User
+from django.core.cache import cache
 from django.db import connection
 from django.test import TestCase
 from django.test import override_settings
@@ -18,6 +19,7 @@ from guardian.shortcuts import assign_perm
 from rest_framework import status
 
 from documents.caching import get_llm_suggestion_cache
+from documents.caching import get_suggestion_cache_key
 from documents.caching import set_llm_suggestions_cache
 from documents.models import Correspondent
 from documents.models import Document
@@ -342,7 +344,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         super().setUp()
 
     @patch("documents.views.get_llm_suggestion_cache")
-    @patch("documents.views.refresh_suggestions_cache")
+    @patch("documents.views.refresh_llm_suggestions_cache")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -383,12 +385,15 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(response.json()["tags"], [self.tag1.pk])
         mock_get_cache.assert_called_once_with(
             self.document.pk,
-            backend="mock_backend",
+            backend=f"mock_backend:user={self.user.pk}",
+        )
+        mock_refresh_cache.assert_called_once_with(
+            self.document.pk,
+            backend=f"mock_backend:user={self.user.pk}",
         )
-        mock_refresh_cache.assert_called_once_with(self.document.pk)
 
     @patch("documents.views.get_llm_suggestion_cache")
-    @patch("documents.views.refresh_suggestions_cache")
+    @patch("documents.views.refresh_llm_suggestions_cache")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -524,7 +529,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(
             get_llm_suggestion_cache(
                 self.document.pk,
-                backend="mock_backend:de-de",
+                backend=f"mock_backend:de-de:user={self.user.pk}",
             ).suggestions["title"],
             "KI Title",
         )
@@ -563,7 +568,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(
             get_llm_suggestion_cache(
                 self.document.pk,
-                backend="mock_backend:fr-fr",
+                backend=f"mock_backend:fr-fr:user={self.user.pk}",
             ).suggestions["title"],
             "Titre IA",
         )
@@ -600,7 +605,79 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertIsNotNone(
             get_llm_suggestion_cache(
                 self.document.pk,
-                backend="mock_backend:model-a:http://endpoint-a",
+                backend=(f"mock_backend:model-a:http://endpoint-a:user={self.user.pk}"),
+            ),
+        )
+
+    @patch("documents.views.get_ai_document_classification")
+    @override_settings(
+        AI_ENABLED=True,
+        LLM_BACKEND="mock_backend",
+    )
+    def test_ai_suggestions_cache_variants_coexist_per_requesting_user(
+        self,
+        mock_get_ai_classification,
+    ) -> None:
+        """
+        GIVEN:
+            - One user has populated the document's LLM suggestion cache
+            - A second user requests suggestions for the same document and
+              backend
+        WHEN:
+            - The second request is made
+        THEN:
+            - The first user's prompt-derived result is not reused
+            - The classification runs with the second user's visibility
+              context without evicting the first user's result
+        """
+        second_user = User.objects.create_superuser(username="second_user")
+        empty_choices = {
+            "tags": {"existing_ids": [], "new_names": []},
+            "correspondents": {"existing_ids": [], "new_names": []},
+            "document_types": {"existing_ids": [], "new_names": []},
+            "storage_paths": {"existing_ids": [], "new_names": []},
+            "dates": [],
+        }
+        mock_get_ai_classification.side_effect = [
+            {"title": "First user's result", **empty_choices},
+            {"title": "Second user's result", **empty_choices},
+        ]
+
+        self.client.force_login(user=self.user)
+        first_response = self.client.get(
+            f"/api/documents/{self.document.pk}/ai_suggestions/",
+        )
+        self.client.force_login(user=second_user)
+        second_response = self.client.get(
+            f"/api/documents/{self.document.pk}/ai_suggestions/",
+        )
+        self.client.force_login(user=self.user)
+        first_cached_response = self.client.get(
+            f"/api/documents/{self.document.pk}/ai_suggestions/",
+        )
+
+        self.assertEqual(first_response.json()["title"], "First user's result")
+        self.assertEqual(second_response.json()["title"], "Second user's result")
+        self.assertEqual(
+            first_cached_response.json()["title"],
+            "First user's result",
+        )
+        self.assertEqual(mock_get_ai_classification.call_count, 2)
+        mock_get_ai_classification.assert_called_with(
+            self.document,
+            second_user,
+            None,
+        )
+        self.assertIsNotNone(
+            get_llm_suggestion_cache(
+                self.document.pk,
+                backend=f"mock_backend:user={second_user.pk}",
+            ),
+        )
+        self.assertIsNotNone(
+            get_llm_suggestion_cache(
+                self.document.pk,
+                backend=f"mock_backend:user={self.user.pk}",
             ),
         )
 
@@ -786,8 +863,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(response.json()["tags"], [])
         self.assertEqual(response.json()["suggested_tags"], [])
 
-    def test_invalidate_suggestions_cache(self) -> None:
-        self.client.force_login(user=self.user)
+    def test_document_save_invalidates_all_suggestion_caches(self) -> None:
         suggestions = {
             "title": "AI Title",
             "tags": ["tag1", "tag2"],
@@ -796,11 +872,18 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             "storage_paths": ["path1"],
             "dates": ["2023-01-01"],
         }
+        standard_cache_key = get_suggestion_cache_key(self.document.pk)
+        cache.set(standard_cache_key, "classifier suggestions")
         set_llm_suggestions_cache(
             self.document.pk,
             suggestions,
             backend="mock_backend",
         )
+        set_llm_suggestions_cache(
+            self.document.pk,
+            {**suggestions, "title": "Other Variant"},
+            backend="other_backend:user=2",
+        )
         self.assertEqual(
             get_llm_suggestion_cache(
                 self.document.pk,
@@ -808,17 +891,26 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             ).suggestions,
             suggestions,
         )
-        # post_save signal triggered
+        self.assertEqual(cache.get(standard_cache_key), "classifier suggestions")
+
         update_llm_suggestions_cache(
             sender=None,
             instance=self.document,
         )
+
+        self.assertIsNone(cache.get(standard_cache_key))
         self.assertIsNone(
             get_llm_suggestion_cache(
                 self.document.pk,
                 backend="mock_backend",
             ),
         )
+        self.assertIsNone(
+            get_llm_suggestion_cache(
+                self.document.pk,
+                backend="other_backend:user=2",
+            ),
+        )
 
 
 class TestAIChatStreamingView(DirectoriesMixin, TestCase):

src/documents/views.py

@@ -113,6 +113,7 @@ from documents.bulk_download import OriginalsOnlyStrategy
 from documents.caching import get_llm_suggestion_cache
 from documents.caching import get_metadata_cache
 from documents.caching import get_suggestion_cache
+from documents.caching import refresh_llm_suggestions_cache
 from documents.caching import refresh_metadata_cache
 from documents.caching import refresh_suggestions_cache
 from documents.caching import set_llm_suggestions_cache
@@ -1540,6 +1541,7 @@ class DocumentViewSet(
                 ai_config.llm_model,
                 ai_config.llm_endpoint,
                 output_language,
+                f"user={request.user.pk}",
             )
             if part
         )
@@ -1555,8 +1557,11 @@ class DocumentViewSet(
             # freshly for this requester on every request, cache hit or not,
             # so a resolved id cached for one user's visibility can never be
             # handed unfiltered to a second, less-privileged requester of
-            # the same (backend-keyed, not user-keyed) cache entry.
-            refresh_suggestions_cache(doc.pk)
+            # the same (backend + user-keyed) cache entry.
+            refresh_llm_suggestions_cache(
+                doc.pk,
+                backend=llm_cache_backend,
+            )
             llm_suggestions = cached_llm_suggestions.suggestions
         else:
             try: