fix: propagate Xinference rerank failures instead of returning empty results (#10051)

The Xinference rerank provider swallowed upstream failures and an
uninitialized model by returning an empty list, which the retrieval
manager read as a legitimate empty rerank and used to overwrite the
fused candidates with nothing. Other providers (e.g. TEI) raise, and
the manager only preserves the fused results on that path, so a
transient Xinference outage silently degraded every knowledge-base
query to zero hits. Raise instead, matching the provider contract and
letting the manager's degradation keep the unreranked results.

Fixes #10000

astrbot/core/provider/sources/xinference_rerank_source.py

@@ -91,8 +91,7 @@ class XinferenceRerankProvider(RerankProvider):
         top_n: int | None = None,
     ) -> list[RerankResult]:
         if not self.model:
-            logger.error("Xinference rerank model is not initialized.")
-            return []
+            raise RuntimeError("Xinference rerank model is not initialized")
         try:
             response = await self.model.rerank(documents, query, top_n)
             results = response.get("results", [])
@@ -111,9 +110,12 @@ class XinferenceRerankProvider(RerankProvider):
                 for result in results
             ]
         except Exception as e:
+            # An empty list reads as a legitimate empty rerank to the caller,
+            # which would overwrite the fused candidates with nothing (#10000).
+            # Propagate so the retrieval manager keeps the unreranked results.
             logger.error(f"Xinference rerank failed: {e}")
             logger.debug(f"Xinference rerank failed with exception: {e}", exc_info=True)
-            return []
+            raise
 
     async def terminate(self) -> None:
         """关闭客户端会话"""

tests/test_xinference_rerank_source.py

@@ -0,0 +1,38 @@
+"""Regression tests for the Xinference rerank provider failure contract (#10000).
+
+The provider used to swallow upstream failures and return an empty list,
+which the retrieval manager read as a legitimate empty rerank result and
+used to overwrite the fused candidates. Failures now propagate so the
+manager can fall back to the unreranked results.
+"""
+
+from unittest.mock import AsyncMock, Mock
+
+import pytest
+
+from astrbot.core.provider.sources.xinference_rerank_source import (
+    XinferenceRerankProvider,
+)
+
+
+@pytest.fixture
+def provider() -> XinferenceRerankProvider:
+    instance = XinferenceRerankProvider.__new__(XinferenceRerankProvider)
+    instance.model = Mock()
+    return instance
+
+
+@pytest.mark.asyncio
+async def test_rerank_failure_propagates_instead_of_empty_list(provider):
+    provider.model.rerank = AsyncMock(side_effect=RuntimeError("upstream down"))
+
+    with pytest.raises(RuntimeError, match="upstream down"):
+        await provider.rerank(query="q", documents=["a", "b"])
+
+
+@pytest.mark.asyncio
+async def test_uninitialized_model_raises_instead_of_empty_list(provider):
+    provider.model = None
+
+    with pytest.raises(RuntimeError, match="not initialized"):
+        await provider.rerank(query="q", documents=["a"])