add the fix to streaming query as well along with UTs

Author: Radovan Fuchs

src/utils/endpoints.py

@@ -671,11 +671,19 @@ async def cleanup_after_streaming(
             session.query(UserConversation).filter_by(id=conversation_id).first()
         )
         if not existing_conversation:
-            topic_summary = await get_topic_summary_func(
-                query_request.query,
-                client,
-                llama_stack_model_id,
-            )
+            # Check if topic summary should be generated (default: True)
+            should_generate = query_request.generate_topic_summary
+
+            if should_generate:
+                logger.debug("Generating topic summary for new conversation")
+                topic_summary = await get_topic_summary_func(
+                    query_request.query, client, llama_stack_model_id
+                )
+            else:
+                logger.debug(
+                    "Topic summary generation disabled by request parameter"
+                )
+                topic_summary = None
 
     completed_at = datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ")
 

tests/unit/utils/test_endpoints.py

@@ -1036,3 +1036,101 @@ def test_create_referenced_documents_invalid_urls(self) -> None:
         assert result[0].doc_title == "not-a-valid-url"
         assert result[1].doc_url == AnyUrl("https://example.com/doc1")
         assert result[1].doc_title == "doc1"
+
+
+@pytest.mark.asyncio
+async def test_cleanup_after_streaming_generate_topic_summary_default_true(
+    mocker: MockerFixture,
+) -> None:
+    """Test that topic summary is generated by default for new conversations."""
+    mock_is_transcripts_enabled = mocker.Mock(return_value=False)
+    mock_get_topic_summary = mocker.AsyncMock(return_value="Generated topic")
+    mock_store_transcript = mocker.Mock()
+    mock_persist_conversation = mocker.Mock()
+    mock_client = mocker.AsyncMock()
+    mock_config = mocker.Mock()
+
+    mock_session = mocker.Mock()
+    mock_session.query.return_value.filter_by.return_value.first.return_value = None
+    mock_session.__enter__ = mocker.Mock(return_value=mock_session)
+    mock_session.__exit__ = mocker.Mock(return_value=None)
+    mocker.patch("utils.endpoints.get_session", return_value=mock_session)
+
+    mocker.patch("utils.endpoints.create_referenced_documents_with_metadata", return_value=[])
+    mocker.patch("utils.endpoints.store_conversation_into_cache")
+
+    query_request = QueryRequest(query="test query")
+
+    await endpoints.cleanup_after_streaming(
+        user_id="test_user",
+        conversation_id="test_conv_id",
+        model_id="test_model",
+        provider_id="test_provider",
+        llama_stack_model_id="test_llama_model",
+        query_request=query_request,
+        summary=mocker.Mock(llm_response="test response", tool_calls=[]),
+        metadata_map={},
+        started_at="2024-01-01T00:00:00Z",
+        client=mock_client,
+        config=mock_config,
+        skip_userid_check=False,
+        get_topic_summary_func=mock_get_topic_summary,
+        is_transcripts_enabled_func=mock_is_transcripts_enabled,
+        store_transcript_func=mock_store_transcript,
+        persist_user_conversation_details_func=mock_persist_conversation,
+    )
+
+    mock_get_topic_summary.assert_called_once_with(
+        "test query", mock_client, "test_llama_model"
+    )
+
+    mock_persist_conversation.assert_called_once()
+    assert mock_persist_conversation.call_args[1]["topic_summary"] == "Generated topic"
+
+
+@pytest.mark.asyncio
+async def test_cleanup_after_streaming_generate_topic_summary_explicit_false(
+    mocker: MockerFixture,
+) -> None:
+    """Test that topic summary is NOT generated when explicitly set to False."""
+    mock_is_transcripts_enabled = mocker.Mock(return_value=False)
+    mock_get_topic_summary = mocker.AsyncMock(return_value="Generated topic")
+    mock_store_transcript = mocker.Mock()
+    mock_persist_conversation = mocker.Mock()
+    mock_client = mocker.AsyncMock()
+    mock_config = mocker.Mock()
+
+    mock_session = mocker.Mock()
+    mock_session.query.return_value.filter_by.return_value.first.return_value = None
+    mock_session.__enter__ = mocker.Mock(return_value=mock_session)
+    mock_session.__exit__ = mocker.Mock(return_value=None)
+    mocker.patch("utils.endpoints.get_session", return_value=mock_session)
+
+    mocker.patch("utils.endpoints.create_referenced_documents_with_metadata", return_value=[])
+    mocker.patch("utils.endpoints.store_conversation_into_cache")
+
+    query_request = QueryRequest(query="test query", generate_topic_summary=False)
+
+    await endpoints.cleanup_after_streaming(
+        user_id="test_user",
+        conversation_id="test_conv_id",
+        model_id="test_model",
+        provider_id="test_provider",
+        llama_stack_model_id="test_llama_model",
+        query_request=query_request,
+        summary=mocker.Mock(llm_response="test response", tool_calls=[]),
+        metadata_map={},
+        started_at="2024-01-01T00:00:00Z",
+        client=mock_client,
+        config=mock_config,
+        skip_userid_check=False,
+        get_topic_summary_func=mock_get_topic_summary,
+        is_transcripts_enabled_func=mock_is_transcripts_enabled,
+        store_transcript_func=mock_store_transcript,
+        persist_user_conversation_details_func=mock_persist_conversation,
+    )
+
+    mock_get_topic_summary.assert_not_called()
+
+    mock_persist_conversation.assert_called_once()
+    assert mock_persist_conversation.call_args[1]["topic_summary"] is None