feat: FAISS in OpenSearch: check existing index (#3101)

* Add check for mapping for existing indices

* Add test

* Check if "method" field exists

haystack/document_stores/opensearch.py

@@ -443,10 +443,10 @@ class OpenSearchDocumentStore(BaseElasticsearchDocumentStore):
                             "space_type"
                         ]
 
+                    # Check if desired index settings are equal to settings in existing index
                     embedding_field_similarity = self.space_type_to_similarity[embedding_field_space_type]
-                    if embedding_field_similarity == self.similarity:
+                    if embedding_field_similarity != self.similarity:
-                        self.embeddings_field_supports_similarity = True
+                        self.embeddings_field_supports_similarity = False
-                    else:
                         logger.warning(
                             f"Embedding field '{self.embedding_field}' is optimized for similarity '{embedding_field_similarity}'. "
                             f"Falling back to slow exact vector calculation. "
@@ -455,6 +455,22 @@ class OpenSearchDocumentStore(BaseElasticsearchDocumentStore):
                             f"e.g. `OpenSearchDocumentStore(index='my_new_{self.similarity}_index', similarity='{self.similarity}')`."
                         )
 
+                    # Check if desired knn engine is same as engine in existing index
+                    elif (
+                        "method" in mappings["properties"][self.embedding_field]
+                        and mappings["properties"][self.embedding_field]["method"]["engine"] != self.knn_engine
+                    ):
+                        self.embeddings_field_supports_similarity = False
+                        embedding_field_engine = mappings["properties"][self.embedding_field]["method"]["engine"]
+                        logger.warning(
+                            f"Embedding field '{self.embedding_field}' was initially created with knn_engine "
+                            f"'{embedding_field_engine}', but knn_engine was set to '{self.knn_engine}' when "
+                            f"initializing OpenSearchDocumentStore. "
+                            f"Falling back to slow exact vector calculation."
+                        )
+                    else:
+                        self.embeddings_field_supports_similarity = True
+
                 # Adjust global ef_search setting (nmslib only). If not set, default is 512.
                 ef_search = index_settings.get("knn.algo_param", {"ef_search": 512}).get("ef_search", 512)
                 if self.index_type == "hnsw" and ef_search != 20:

test/document_stores/test_opensearch.py

@@ -206,6 +206,20 @@ class TestOpenSearchDocumentStore:
                 # docs with an original embedding should have the new one
                 assert cloned_field_name in meta
 
+    @pytest.mark.integration
+    def test_change_knn_engine(self, ds, caplog):
+        assert ds.embeddings_field_supports_similarity == True
+        index_name = ds.index
+        with caplog.at_level(logging.WARNING):
+            ds = OpenSearchDocumentStore(port=9201, knn_engine="faiss", index=index_name)
+            warning = (
+                "Embedding field 'embedding' was initially created with knn_engine 'nmslib', but knn_engine was "
+                "set to 'faiss' when initializing OpenSearchDocumentStore. Falling back to slow exact vector "
+                "calculation."
+            )
+            assert ds.embeddings_field_supports_similarity == False
+            assert warning in caplog.text
+
     # Unit tests
 
     @pytest.mark.unit