Enhancement: prevent duplication of llm suggestion requests

src/documents/caching.py

@@ -2,20 +2,25 @@ from __future__ import annotations
 
 import logging
 import pickle
+import time
 from binascii import hexlify
 from collections import OrderedDict
 from dataclasses import dataclass
+from hashlib import sha256
 from typing import TYPE_CHECKING
 from typing import Any
 from typing import Final
+from uuid import uuid4
 
 from django.conf import settings
 from django.core.cache import cache
 from django.core.cache import caches
 
 from documents.models import Document
+from paperless_ai.ai_classifier import get_ai_document_classification
 
 if TYPE_CHECKING:
+    from django.contrib.auth.models import User
     from django.core.cache.backends.base import BaseCache
 
     from documents.classifier import DocumentClassifier
@@ -223,6 +228,56 @@ def get_llm_suggestion_cache(
     return None
 
 
+def retrieve_llm_suggestions(
+    document: Document,
+    user: User | None,
+    output_language: str | None,
+    backend: str,
+    *,
+    lock_timeout: int,
+) -> dict:
+    """Return cached LLM suggestions, generating them once across workers."""
+    lock_key = (
+        f"{get_suggestion_cache_key(document.pk)}_llm_lock_"
+        f"{sha256(backend.encode()).hexdigest()}"
+    )
+
+    while True:
+        cached = get_llm_suggestion_cache(document.pk, backend)
+        if cached is not None:
+            refresh_suggestions_cache(document.pk)
+            return cached.suggestions
+
+        lock_token = uuid4().hex
+        if cache.add(lock_key, lock_token, lock_timeout):
+            try:
+                # The cache may have been populated while acquiring the lock.
+                cached = get_llm_suggestion_cache(document.pk, backend)
+                if cached is not None:
+                    refresh_suggestions_cache(document.pk)
+                    return cached.suggestions
+
+                suggestions = get_ai_document_classification(
+                    document,
+                    user,
+                    output_language,
+                )
+                set_llm_suggestions_cache(
+                    document.pk,
+                    suggestions,
+                    backend=backend,
+                )
+                return suggestions
+            finally:
+                # Do not remove a replacement lock if this one expired while
+                # generation was still running.
+                if cache.get(lock_key) == lock_token:
+                    cache.delete(lock_key)
+
+        # Another worker is generating suggestions, poll to avoid another LLM request
+        time.sleep(0.1)
+
+
 def set_llm_suggestions_cache(
     document_id: int,
     suggestions: dict,

src/documents/tests/test_api_documents.py

@@ -2486,7 +2486,7 @@ class TestDocumentApi(DirectoriesMixin, ConsumeTaskMixin, APITestCase):
         response = self.client.get("/api/documents/34676/suggestions/")
         self.assertEqual(response.status_code, status.HTTP_404_NOT_FOUND)
 
-    @mock.patch("documents.views.get_ai_document_classification")
+    @mock.patch("documents.caching.get_ai_document_classification")
     @override_settings(AI_ENABLED=True)
     def test_suggestions_still_uses_classifier_when_ai_enabled(
         self,

src/documents/tests/test_caching.py

@@ -1,6 +1,12 @@
 import pickle
+from concurrent.futures import ThreadPoolExecutor
+from threading import Event
+from threading import Lock
+
+from django.core.cache import cache
 
 from documents.caching import StoredLRUCache
+from documents.caching import retrieve_llm_suggestions
 
 
 def test_lru_cache_entries() -> None:
@@ -43,3 +49,60 @@ def test_stored_lru_cache_key_ttl(mocker) -> None:
     assert key == "test_key"
     assert timeout == 321
     assert pickle.loads(data) == {"x": "X", "y": "Y"}
+
+
+def test_llm_suggestions_are_generated_once_for_concurrent_requests(mocker) -> None:
+    cache.clear()
+    generation_started = Event()
+    finish_generation = Event()
+    waiter_started = Event()
+    call_lock = Lock()
+    calls = 0
+    suggestions = {"title": "Generated once"}
+    document = mocker.Mock(pk=42)
+    user = mocker.Mock()
+
+    def generate(*args) -> dict:
+        nonlocal calls
+        with call_lock:
+            calls += 1
+        generation_started.set()
+        assert finish_generation.wait(timeout=2)
+        return suggestions
+
+    def wait_for_generation(_interval: float) -> None:
+        waiter_started.set()
+        assert finish_generation.wait(timeout=2)
+
+    mock_get_classification = mocker.patch(
+        "documents.caching.get_ai_document_classification",
+        side_effect=generate,
+    )
+    mocker.patch("documents.caching.time.sleep", side_effect=wait_for_generation)
+
+    with ThreadPoolExecutor(max_workers=2) as executor:
+        first = executor.submit(
+            retrieve_llm_suggestions,
+            document,
+            user,
+            None,
+            "ollama:model",
+            lock_timeout=10,
+        )
+        assert generation_started.wait(timeout=2)
+        second = executor.submit(
+            retrieve_llm_suggestions,
+            document,
+            user,
+            None,
+            "ollama:model",
+            lock_timeout=10,
+        )
+        assert waiter_started.wait(timeout=2)
+        finish_generation.set()
+
+        assert first.result(timeout=2) == suggestions
+        assert second.result(timeout=2) == suggestions
+
+    assert calls == 1
+    mock_get_classification.assert_called_once_with(document, user, None)

src/documents/tests/test_views.py

@@ -441,7 +441,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(response.json()["tags"], [])
         self.assertEqual(response.json()["suggested_tags"], [])
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -491,7 +491,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             None,
         )
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -529,7 +529,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             "KI Title",
         )
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -568,7 +568,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             "Titre IA",
         )
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -604,7 +604,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             ),
         )
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="openai-like",
@@ -633,7 +633,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             get_llm_suggestion_cache(self.document.pk, backend="openai-like"),
         )
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="openai-like",
@@ -660,7 +660,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             get_llm_suggestion_cache(self.document.pk, backend="openai-like"),
         )
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -698,7 +698,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(response.json()["tags"], [self.tag1.pk])
         self.assertEqual(response.json()["suggested_tags"], ["Follow-up"])
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",
@@ -737,7 +737,7 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
         self.assertEqual(response.json()["tags"], [self.tag1.pk])
         self.assertEqual(response.json()["suggested_tags"], [])
 
-    @patch("documents.views.get_ai_document_classification")
+    @patch("documents.caching.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,
         LLM_BACKEND="mock_backend",

src/documents/views.py

@@ -115,7 +115,7 @@ from documents.caching import get_metadata_cache
 from documents.caching import get_suggestion_cache
 from documents.caching import refresh_metadata_cache
 from documents.caching import refresh_suggestions_cache
-from documents.caching import set_llm_suggestions_cache
+from documents.caching import retrieve_llm_suggestions
 from documents.caching import set_metadata_cache
 from documents.caching import set_suggestions_cache
 from documents.classifier import load_classifier
@@ -246,7 +246,6 @@ from paperless.parsers.remote import RemoteEngineConfig
 from paperless.serialisers import GroupSerializer
 from paperless.serialisers import UserSerializer
 from paperless.views import StandardPagination
-from paperless_ai.ai_classifier import get_ai_document_classification
 from paperless_ai.ai_classifier import get_llm_output_language
 from paperless_ai.chat import stream_chat_with_documents
 from paperless_ai.exceptions import LLMTimeoutError
@@ -1560,10 +1559,13 @@ class DocumentViewSet(
             llm_suggestions = cached_llm_suggestions.suggestions
         else:
             try:
-                llm_suggestions = get_ai_document_classification(
+                llm_suggestions = retrieve_llm_suggestions(
                     doc,
                     request.user,
                     output_language,
+                    llm_cache_backend,
+                    # Classification, localization + 30s
+                    lock_timeout=(2 * ai_config.llm_request_timeout) + 30,
                 )
             except ValueError as exc:
                 logger.exception(
@@ -1588,11 +1590,6 @@ class DocumentViewSet(
                     {"ai": [_("AI backend request timed out.")]},
                     status=status.HTTP_503_SERVICE_UNAVAILABLE,
                 )
-            set_llm_suggestions_cache(
-                doc.pk,
-                llm_suggestions,
-                backend=llm_cache_backend,
-            )
 
         tags_choice: TaxonomyChoiceDict = llm_suggestions["tags"]
         correspondents_choice: TaxonomyChoiceDict = llm_suggestions["correspondents"]