Fix: better LLM errors

src/documents/tests/test_views.py

@@ -32,6 +32,7 @@ from documents.signals.handlers import update_llm_suggestions_cache
 from documents.tests.utils import DirectoriesMixin
 from documents.tests.utils import read_streaming_response
 from paperless.models import ApplicationConfiguration
+from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 
 
@@ -737,6 +738,38 @@ class TestAISuggestions(DirectoriesMixin, TestCase):
             get_llm_suggestion_cache(self.document.pk, backend="openai-like"),
         )
 
+    @patch("documents.views.get_ai_document_classification")
+    @override_settings(
+        AI_ENABLED=True,
+        LLM_BACKEND="openai-like",
+    )
+    def test_ai_suggestions_with_llm_provider_error(
+        self,
+        mock_get_ai_classification,
+    ) -> None:
+        mock_get_ai_classification.side_effect = LLMProviderError(
+            "confidential provider response",
+        )
+
+        self.client.force_login(user=self.user)
+        response = self.client.get(
+            f"/api/documents/{self.document.pk}/ai_suggestions/",
+        )
+
+        self.assertEqual(response.status_code, status.HTTP_502_BAD_GATEWAY)
+        self.assertEqual(
+            response.json(),
+            {
+                "ai": [
+                    "AI backend rejected the request. Check logs for details.",
+                ],
+            },
+        )
+        self.assertNotIn("confidential provider response", response.content.decode())
+        self.assertIsNone(
+            get_llm_suggestion_cache(self.document.pk, backend="openai-like"),
+        )
+
     @patch("documents.views.get_ai_document_classification")
     @override_settings(
         AI_ENABLED=True,

src/documents/views.py

@@ -251,6 +251,7 @@ 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 LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 from paperless_ai.matching import extract_unmatched_names
 from paperless_ai.matching import match_correspondents_by_name
@@ -1602,6 +1603,22 @@ class DocumentViewSet(
                     {"ai": [_("AI backend request timed out.")]},
                     status=status.HTTP_503_SERVICE_UNAVAILABLE,
                 )
+            except LLMProviderError:
+                logger.exception(
+                    "AI backend rejected the request for document %s",
+                    doc.pk,
+                )
+                return Response(
+                    {
+                        "ai": [
+                            _(
+                                "AI backend rejected the request. "
+                                "Check logs for details.",
+                            ),
+                        ],
+                    },
+                    status=status.HTTP_502_BAD_GATEWAY,
+                )
             set_llm_suggestions_cache(
                 doc.pk,
                 llm_suggestions,

src/paperless_ai/client.py

@@ -22,6 +22,7 @@ from paperless.network import validate_outbound_http_url
 from paperless_ai.base_model import ClassificationSuggestions
 from paperless_ai.base_model import DocumentClassifierSchema
 from paperless_ai.base_model import model_to_classification_suggestions
+from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 
 logger = logging.getLogger("paperless_ai.client")
@@ -132,7 +133,7 @@ class AIClient:
         from llama_index.core.llms import ChatMessage
 
         if self.settings.llm_backend == LLMBackend.OLLAMA:
-            with self._normalize_timeouts():
+            with self._normalize_errors():
                 result = self.llm.chat(
                     [ChatMessage(role="user", content=prompt)],
                     format=DocumentClassifierSchema.model_json_schema(),
@@ -153,7 +154,7 @@ class AIClient:
             content=f"{prompt}\n\n"
             f"Answer by calling the {tool.metadata.name} tool. Do not write the answer as text.",
         )
-        with self._normalize_timeouts():
+        with self._normalize_errors():
             result = self.llm.chat_with_tools(
                 tools=[tool],
                 user_msg=user_msg,
@@ -173,7 +174,7 @@ class AIClient:
         )
 
     @contextmanager
-    def _normalize_timeouts(self) -> Iterator[None]:
+    def _normalize_errors(self) -> Iterator[None]:
         try:
             yield
         except httpx.TimeoutException as exc:
@@ -181,8 +182,18 @@ class AIClient:
         except Exception as exc:
             if self._is_openai_timeout(exc):
                 raise LLMTimeoutError from exc
+            if self._is_openai_status_error(exc):
+                raise LLMProviderError from exc
             raise
 
+    def _is_openai_status_error(self, exc: Exception) -> bool:
+        if self.settings.llm_backend != LLMBackend.OPENAI_LIKE:
+            return False
+
+        from openai import APIStatusError
+
+        return isinstance(exc, APIStatusError)
+
     def _is_openai_timeout(self, exc: Exception) -> bool:
         if self.settings.llm_backend != LLMBackend.OPENAI_LIKE:
             return False

src/paperless_ai/exceptions.py

@@ -1,2 +1,6 @@
 class LLMTimeoutError(Exception):
     pass
+
+
+class LLMProviderError(Exception):
+    """The LLM backend rejected the request."""

src/paperless_ai/tests/test_client.py

@@ -11,6 +11,7 @@ from llama_index.core.llms.llm import ToolSelection
 from paperless_ai.client import LLM_SYSTEM_PROMPT
 from paperless_ai.client import PLACEHOLDER_API_KEY
 from paperless_ai.client import AIClient
+from paperless_ai.exceptions import LLMProviderError
 from paperless_ai.exceptions import LLMTimeoutError
 
 
@@ -214,6 +215,30 @@ def test_run_llm_query_openai_timeout_raises_local_error(
         client.run_llm_query("test_prompt")
 
 
+def test_run_llm_query_openai_status_error_raises_provider_error(
+    mock_ai_config,
+    mock_openai_llm,
+):
+    mock_ai_config.llm_backend = "openai-like"
+    mock_ai_config.llm_model = "test_model"
+    mock_ai_config.llm_endpoint = "http://test-url"
+
+    request = httpx.Request("POST", "http://test-url/v1/chat/completions")
+    body = {"error": {"message": "Thinking mode does not support this tool_choice"}}
+    mock_openai_llm.return_value.chat_with_tools.side_effect = openai.BadRequestError(
+        "Error code: 400",
+        response=httpx.Response(400, request=request, json=body),
+        body=body,
+    )
+
+    client = AIClient()
+
+    with pytest.raises(LLMProviderError) as exc_info:
+        client.run_llm_query("test_prompt")
+    assert str(exc_info.value) == ""
+    assert isinstance(exc_info.value.__cause__, openai.BadRequestError)
+
+
 def test_run_llm_query_httpx_timeout_raises_local_error(
     mock_ai_config,
     mock_ollama_llm,