fix(rag): detect ChromaDB dim mismatch and recreate collection on provider switch

tests/test_rag_engine.py

@@ -52,7 +52,88 @@ def test_rag_engine_chroma(mock_get_chroma, mock_embed):
   results = engine.search("hello", top_k=1)
   assert len(results) == 1
   assert results[0]["id"] == "doc1"
-  assert results[0]["document"] == "hello world"
+  engine.delete_documents(["doc1"])
+  mock_collection.delete.assert_called_once_with(ids=["doc1"])
+
+@patch('src.rag_engine.LocalEmbeddingProvider.embed')
+@patch('src.rag_engine._get_chromadb')
+def test_rag_collection_dim_mismatch_recreates_collection(mock_get_chroma, mock_embed):
+ """
+ Regression test for the live_gui_test_hardening_v2 followup
+ (RAG dimension-mismatch flake in test_rag_phase4_stress).
+
+ Scenario: a ChromaDB collection exists on disk with vectors from a
+ previous embedding provider (e.g. Gemini, 3072-dim), but the current
+ config uses a different provider (e.g. local SentenceTransformers,
+ 384-dim). Without the dim check, upsert silently corrupts the
+ collection and search() later fails with
+ "Collection expecting embedding with dimension of 3072, got 384".
+
+ Expected: RAGEngine.__init__ detects the mismatch, deletes the
+ mismatched collection, and recreates it empty so subsequent indexing
+ uses the correct dim.
+ """
+ mock_chroma = MagicMock()
+ mock_settings = MagicMock()
+ mock_get_chroma.return_value = (mock_chroma, mock_settings)
+
+ mock_embed.return_value = [[0.1] * 384]
+ mock_collection = MagicMock()
+ mock_collection.get.return_value = {
+  "embeddings": [[0.1] * 3072],
+  "metadatas":   [{}],
+  "ids":         ["stale_doc_1"],
+ }
+ mock_collection.name = "test"
+
+ mock_client = MagicMock()
+ mock_client.get_or_create_collection.return_value = mock_collection
+ mock_chroma.PersistentClient.return_value = mock_client
+
+ vs_config = models.VectorStoreConfig(provider='chroma', collection_name='test')
+ config = models.RAGConfig(enabled=True, vector_store=vs_config, embedding_provider='local')
+
+ with patch('src.rag_engine._get_sentence_transformers') as mock_st:
+  mock_st.return_value = MagicMock()
+  engine = RAGEngine(config)
+  assert engine.collection == mock_collection
+  mock_client.delete_collection.assert_called_once_with("test")
+  assert mock_client.get_or_create_collection.call_count == 2
+
+@patch('src.rag_engine.LocalEmbeddingProvider.embed')
+@patch('src.rag_engine._get_chromadb')
+def test_rag_collection_dim_match_preserves_collection(mock_get_chroma, mock_embed):
+ """
+ Companion test: when the collection's existing dim matches the current
+ provider's dim, the engine must NOT delete the collection (which would
+ discard indexed data).
+ """
+ mock_chroma = MagicMock()
+ mock_settings = MagicMock()
+ mock_get_chroma.return_value = (mock_chroma, mock_settings)
+
+ mock_embed.return_value = [[0.1] * 384]
+ mock_collection = MagicMock()
+ mock_collection.get.return_value = {
+  "embeddings": [[0.1] * 384],
+  "metadatas":   [{"path": "file_25.txt"}],
+  "ids":         ["doc_25_0"],
+ }
+ mock_collection.name = "test"
+
+ mock_client = MagicMock()
+ mock_client.get_or_create_collection.return_value = mock_collection
+ mock_chroma.PersistentClient.return_value = mock_client
+
+ vs_config = models.VectorStoreConfig(provider='chroma', collection_name='test')
+ config = models.RAGConfig(enabled=True, vector_store=vs_config, embedding_provider='local')
+
+ with patch('src.rag_engine._get_sentence_transformers') as mock_st:
+  mock_st.return_value = MagicMock()
+  engine = RAGEngine(config)
+  assert engine.collection == mock_collection
+  mock_client.delete_collection.assert_not_called()
+  assert mock_client.get_or_create_collection.call_count == 1
 
   engine.delete_documents(["doc1"])
   mock_collection.delete.assert_called_once_with(ids=["doc1"])