[fix] plugin: repeat the request on 5xx as well

A server that is not ready does not always take its time about saying so.
Ollama answers 5xx immediately while a model is still being loaded, and a
gateway with no upstream yet does the same, so the failure arrives at
once rather than as a timeout.

Both mean "not right now" and are worth one more attempt.  4xx keeps
being final -- a wrong API key or an unknown model name is not going to
change between two requests.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

searx/plugins/ai_summary.py

@@ -254,10 +254,13 @@ def _open_upstream(client: httpx.Client, payload: dict[str, t.Any]) -> tuple[t.A
     """Start the streaming completion on the LLM server.
 
     Returns the (already entered) stream context and the response, or
-    ``(None, None)`` if no usable response was received.  A server that does not
-    answer in time is asked again (:py:obj:`UPSTREAM_RETRIES`); a server that
-    *did* answer, but with an error status, is not -- a wrong API key or an
-    unknown model name does not become right when asked twice."""
+    ``(None, None)`` if no usable response was received.
+
+    The request is repeated (:py:obj:`UPSTREAM_RETRIES`) when the server did not
+    answer in time, and when it answered ``5xx`` -- both mean *not right now*,
+    and the most common reason is a model that is still being loaded.  A ``4xx``
+    is not repeated: a wrong API key or an unknown model name does not become
+    right when asked twice."""
 
     for attempt in range(UPSTREAM_RETRIES + 1):
         stream_ctx = client.stream("POST", "/chat/completions", json=payload)
@@ -284,8 +287,15 @@ def _open_upstream(client: httpx.Client, payload: dict[str, t.Any]) -> tuple[t.A
             detail = resp.text.strip()[:200]
         except (httpx.HTTPError, UnicodeDecodeError):  # pragma: no cover
             pass
-        log.warning("LLM server responded with HTTP %s %s", resp.status_code, detail)
         stream_ctx.__exit__(None, None, None)
+
+        # 5xx is the server saying it is not able to answer *right now* -- a
+        # model still loading, a gateway with no upstream yet.  4xx is the
+        # server saying the request is wrong, which a second one would be too.
+        if resp.status_code >= 500 and attempt < UPSTREAM_RETRIES:
+            log.debug("LLM server replied HTTP %s (%s), asking again", resp.status_code, detail)
+            continue
+        log.warning("LLM server responded with HTTP %s %s", resp.status_code, detail)
         return None, None
 
     return None, None  # pragma: no cover - the loop always returns

tests/unit/test_plugin_ai_summary.py

@@ -473,7 +473,7 @@ class PluginAISummary(SearxTestCase):
         self.assertEqual(res.status_code, 502)
         self.assertEqual(2, len(attempts))
 
-    def test_endpoint_does_not_retry_an_error_status(self):
+    def test_endpoint_does_not_retry_a_client_error(self):
         # a wrong API key or an unknown model does not fix itself
         attempts = []
         bad = sse_stream_mock([], status_code=401)
@@ -492,6 +492,30 @@ class PluginAISummary(SearxTestCase):
         self.assertEqual(res.status_code, 502)
         self.assertEqual(1, len(attempts))
 
+    def test_endpoint_retries_a_server_error(self):
+        # 5xx means "not right now" -- e.g. a model that is still loading;
+        # this is the failure that arrives immediately instead of timing out
+        attempts = []
+        ok = sse_stream_mock([{"choices": [{"delta": {"content": "hi"}}]}])
+        bad = sse_stream_mock([], status_code=503)
+
+        @contextmanager
+        def stream(*_args, **_kwargs):
+            attempts.append(1)
+            src = bad if len(attempts) == 1 else ok
+            with src.stream() as resp:
+                yield resp
+
+        client_mock = Mock()
+        client_mock.stream = stream
+        self.mock_upstream(client_mock)
+
+        res = self.client.post("/ai_summary", json={"messages": [{"role": "user", "content": "hi"}]})
+        self.assertEqual(res.status_code, 200)
+        self.assertEqual(2, len(attempts))
+        lines = [json.loads(line) for line in res.data.decode().splitlines() if line]
+        self.assertEqual(lines[0], {"delta": "hi"})
+
     def test_endpoint_upstream_error(self):
         self.mock_upstream(sse_stream_mock([], status_code=500))