From 0656365178cbf242e872433ea153d2837a966aa3 Mon Sep 17 00:00:00 2001 From: Derek Xu Date: Tue, 27 Jan 2026 00:36:49 -0800 Subject: [PATCH 1/2] langfuse degrade --- eval_protocol/adapters/langfuse.py | 6 +++++- tests/chinook/langfuse/generate_traces.py | 4 ++-- tests/chinook/langfuse/test_langfuse_chinook.py | 4 ++-- 3 files changed, 9 insertions(+), 5 deletions(-) diff --git a/eval_protocol/adapters/langfuse.py b/eval_protocol/adapters/langfuse.py index 0174922f..048e695a 100644 --- a/eval_protocol/adapters/langfuse.py +++ b/eval_protocol/adapters/langfuse.py @@ -46,9 +46,13 @@ def __call__( try: - from langfuse import get_client # pyright: ignore[reportPrivateImportUsage] + from langfuse import Langfuse LANGFUSE_AVAILABLE = True + + def get_client(): + """Compatibility shim for langfuse 2.x (returns Langfuse instance).""" + return Langfuse() except ImportError: LANGFUSE_AVAILABLE = False diff --git a/tests/chinook/langfuse/generate_traces.py b/tests/chinook/langfuse/generate_traces.py index 6145b6af..c63e17c8 100644 --- a/tests/chinook/langfuse/generate_traces.py +++ b/tests/chinook/langfuse/generate_traces.py @@ -11,12 +11,12 @@ from tests.chinook.dataset import collect_dataset try: - from langfuse import get_client, observe # pyright: ignore[reportPrivateImportUsage] + from langfuse import Langfuse, observe from pydantic_ai.agent import Agent from pydantic_ai.models.openai import OpenAIChatModel LANGFUSE_AVAILABLE = True - langfuse_client = get_client() + langfuse_client = Langfuse() Agent.instrument_all() diff --git a/tests/chinook/langfuse/test_langfuse_chinook.py b/tests/chinook/langfuse/test_langfuse_chinook.py index edb1eaa7..9be8f65d 100644 --- a/tests/chinook/langfuse/test_langfuse_chinook.py +++ b/tests/chinook/langfuse/test_langfuse_chinook.py @@ -24,10 +24,10 @@ # Langfuse client setup try: - from langfuse import get_client # pyright: ignore[reportPrivateImportUsage] + from langfuse import Langfuse LANGFUSE_AVAILABLE = True - langfuse = get_client() + langfuse = Langfuse() except ImportError: LANGFUSE_AVAILABLE = False langfuse = None From 3f174a71e7d1cbb8720f8ea2b4407d1c9a148385 Mon Sep 17 00:00:00 2001 From: Derek Xu Date: Tue, 27 Jan 2026 00:41:31 -0800 Subject: [PATCH 2/2] skip tests because dataset is gone --- tests/test_adapters_e2e.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/tests/test_adapters_e2e.py b/tests/test_adapters_e2e.py index 2c105315..e62b84e6 100644 --- a/tests/test_adapters_e2e.py +++ b/tests/test_adapters_e2e.py @@ -205,6 +205,7 @@ def test_langfuse_conversation_analysis(self): class TestHuggingFaceAdapterE2E: """End-to-end tests for HuggingFace adapter with real datasets.""" + @pytest.mark.skip(reason="gsm8k dataset no longer available on HuggingFace Hub") def test_gsm8k_adapter_real_data(self): """Test loading real GSM8K data and converting to EvaluationRow.""" try: @@ -318,6 +319,7 @@ def math_transform(row: Dict[str, Any]) -> Dict[str, Any]: print(f" Row {i}: Type={dataset_info.get('type')}, Level={dataset_info.get('level')}") + @pytest.mark.skip(reason="squad dataset no longer available on HuggingFace Hub") def test_custom_dataset_transform(self): """Test adapter with a completely custom transformation.""" try: @@ -663,6 +665,7 @@ def google_books_transform(row: Dict[str, Any]) -> Dict[str, Any]: assert doc_freq > 5, f"Row {i} should have document frequency > 5" +@pytest.mark.skip(reason="gsm8k dataset no longer available on HuggingFace Hub") def test_adapters_integration(): """Test that adapters work with evaluation pipeline.""" print("Testing adapter integration with evaluation pipeline...")