~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.
Python academiaanthropicarxivbravedeep-researchencryptionhome-automationhomeserverlocallocal-deep-researchlocal-llmmistralollamaopenaipubmedresearchresearch-toolretrieval-augmented-generationsearxngself-hosted
fix(library): don't quarantine healthy FAISS index when embedding provider is down (#4512)
LearningCircuit6762ccfaf523
diff --git a/changelog.d/4512.bugfix.md b/changelog.d/4512.bugfix.mdnew file mode 100644index 0000000000..72cb19b946--- /dev/null+++ b/changelog.d/4512.bugfix.md@@ -0,0 +1 @@+A temporarily unreachable embedding provider (e.g. Ollama not running) no longer quarantines a healthy collection index and silently replaces it with an empty one; the search now fails with a clear provider error and the index survives untouched.diff --git a/src/local_deep_research/research_library/services/library_rag_service.py b/src/local_deep_research/research_library/services/library_rag_service.pyindex 07509629ab..a15948006d 100644--- a/src/local_deep_research/research_library/services/library_rag_service.py+++ b/src/local_deep_research/research_library/services/library_rag_service.py@@ -722,15 +722,22 @@ class LibraryRAGService:)self._quarantine_corrupt_index(index_path, reason)else:- try:- # Check for embedding dimension mismatch before loading- current_dim = len(- self.embedding_manager.embeddings.embed_query(- "dimension_check"- )+ # Probe the embedding model OUTSIDE the quarantine+ # try-block: embed_query fails when the embedding+ # provider is unreachable (e.g. Ollama down), which+ # says nothing about the index files. Inside the try+ # it would quarantine a healthy index and silently+ # replace it with an empty one; here it propagates a+ # clear provider error instead.+ current_dim = len(+ self.embedding_manager.embeddings.embed_query(+ "dimension_check")- stored_dim = rag_index.embedding_dimension+ )+ stored_dim = rag_index.embedding_dimension+ try:+ # Check for embedding dimension mismatch before loadingif stored_dim and current_dim != stored_dim:logger.warning(f"Embedding dimension mismatch detected! "diff --git a/tests/research_library/services/test_library_rag_service_coverage.py b/tests/research_library/services/test_library_rag_service_coverage.pyindex e90f37138f..2826452e08 100644--- a/tests/research_library/services/test_library_rag_service_coverage.py+++ b/tests/research_library/services/test_library_rag_service_coverage.py@@ -550,6 +550,45 @@ class TestLoadOrCreateFaissIndex:# Should fall through and create new indexmock_faiss_cls.assert_called_once()+ @patch(f"{_MOD}.FAISS")+ @patch(f"{_MOD}.InMemoryDocstore")+ @patch(f"{_MOD}.IndexFlatIP")+ def test_embedding_provider_failure_raises_and_preserves_index(+ self, mock_flat_ip, mock_docstore, mock_faiss_cls, tmp_path+ ):+ """When the embedding provider is unreachable (e.g. Ollama down),+ the dimension-check embed_query raises. That says nothing about+ the index files, so the error must propagate WITHOUT quarantining+ the healthy index or replacing it with an empty one.+ """+ svc = _make_service()+ svc.embedding_manager = MagicMock()+ svc.embedding_manager.embeddings.embed_query.side_effect = (+ ConnectionError("Ollama connection refused")+ )+ svc.integrity_manager = MagicMock()+ svc.integrity_manager.verify_file.return_value = (True, None)++ idx_path = tmp_path / "healthy.faiss"+ pkl_path = tmp_path / "healthy.pkl"+ idx_path.write_bytes(b"faiss-bytes")+ pkl_path.write_bytes(b"pkl-bytes")++ rag_idx = self._make_rag_index(dim=384)+ rag_idx.index_path = str(idx_path)+ self._patch_get_or_create(svc, rag_idx)++ with pytest.raises(ConnectionError, match="Ollama"):+ svc.load_or_create_faiss_index("coll-1")++ # The healthy index files must be untouched — not quarantined+ assert idx_path.read_bytes() == b"faiss-bytes"+ assert pkl_path.read_bytes() == b"pkl-bytes"+ assert list(tmp_path.glob("*.corrupt-*")) == []+ # And no empty replacement index was created+ mock_faiss_cls.load_local.assert_not_called()+ mock_faiss_cls.assert_not_called()+@patch(f"{_MOD}.FAISS")@patch(f"{_MOD}.InMemoryDocstore")@patch(f"{_MOD}.IndexFlatIP")