import base64 import inspect import json from typing import Any, List, Literal, Optional, cast from unittest.mock import MagicMock import httpx import pytest from langchain_core._api import warn_deprecated from langchain_core.callbacks import BaseCallbackHandler from langchain_core.language_models import BaseChatModel, GenericFakeChatModel from langchain_core.messages import ( AIMessage, AIMessageChunk, BaseMessage, BaseMessageChunk, HumanMessage, SystemMessage, ToolMessage, ) from langchain_core.output_parsers import StrOutputParser from langchain_core.prompts import ChatPromptTemplate from langchain_core.tools import BaseTool, tool from langchain_core.utils.function_calling import ( convert_to_json_schema, tool_example_to_messages, ) from pydantic import BaseModel, Field from pydantic.v1 import BaseModel as BaseModelV1 from pydantic.v1 import Field as FieldV1 from pytest_benchmark.fixture import BenchmarkFixture # type: ignore[import-untyped] from typing_extensions import Annotated, TypedDict from vcr.cassette import Cassette from langchain_tests.unit_tests.chat_models import ( ChatModelTests, ) from langchain_tests.utils.pydantic import PYDANTIC_MAJOR_VERSION def _get_joke_class( schema_type: Literal["pydantic", "typeddict", "json_schema"], ) -> Any: """ :private: """ class Joke(BaseModel): """Joke to tell user.""" setup: str = Field(description="question to set up a joke") punchline: str = Field(description="answer to resolve the joke") def validate_joke(result: Any) -> bool: return isinstance(result, Joke) class JokeDict(TypedDict): """Joke to tell user.""" setup: Annotated[str, ..., "question to set up a joke"] punchline: Annotated[str, ..., "answer to resolve the joke"] def validate_joke_dict(result: Any) -> bool: return all(key in ["setup", "punchline"] for key in result.keys()) if schema_type == "pydantic": return Joke, validate_joke elif schema_type == "typeddict": return JokeDict, validate_joke_dict elif schema_type == "json_schema": return Joke.model_json_schema(), validate_joke_dict else: raise ValueError("Invalid schema type") class _TestCallbackHandler(BaseCallbackHandler): options: list[Optional[dict]] def __init__(self) -> None: super().__init__() self.options = [] def on_chat_model_start( self, serialized: Any, messages: Any, *, options: Optional[dict[str, Any]] = None, **kwargs: Any, ) -> None: self.options.append(options) class _MagicFunctionSchema(BaseModel): input: int = Field(..., gt=-1000, lt=1000) @tool(args_schema=_MagicFunctionSchema) def magic_function(input: int) -> int: """Applies a magic function to an input.""" return input + 2 @tool def magic_function_no_args() -> int: """Calculates a magic function.""" return 5 def _validate_tool_call_message(message: BaseMessage) -> None: assert isinstance(message, AIMessage) assert len(message.tool_calls) == 1 tool_call = message.tool_calls[0] assert tool_call["name"] == "magic_function" assert tool_call["args"] == {"input": 3} assert tool_call["id"] is not None assert tool_call["type"] == "tool_call" def _validate_tool_call_message_no_args(message: BaseMessage) -> None: assert isinstance(message, AIMessage) assert len(message.tool_calls) == 1 tool_call = message.tool_calls[0] assert tool_call["name"] == "magic_function_no_args" assert tool_call["args"] == {} assert tool_call["id"] is not None assert tool_call["type"] == "tool_call" class ChatModelIntegrationTests(ChatModelTests): """Base class for chat model integration tests. Test subclasses must implement the ``chat_model_class`` and ``chat_model_params`` properties to specify what model to test and its initialization parameters. Example: .. code-block:: python from typing import Type from langchain_tests.integration_tests import ChatModelIntegrationTests from my_package.chat_models import MyChatModel class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def chat_model_class(self) -> Type[MyChatModel]: # Return the chat model class to test here return MyChatModel @property def chat_model_params(self) -> dict: # Return initialization parameters for the model. return {"model": "model-001", "temperature": 0} .. note:: API references for individual test methods include troubleshooting tips. Test subclasses must implement the following two properties: chat_model_class The chat model class to test, e.g., ``ChatParrotLink``. Example: .. code-block:: python @property def chat_model_class(self) -> Type[ChatParrotLink]: return ChatParrotLink chat_model_params Initialization parameters for the chat model. Example: .. code-block:: python @property def chat_model_params(self) -> dict: return {"model": "bird-brain-001", "temperature": 0} In addition, test subclasses can control what features are tested (such as tool calling or multi-modality) by selectively overriding the following properties. Expand to see details: .. dropdown:: has_tool_calling Boolean property indicating whether the chat model supports tool calling. By default, this is determined by whether the chat model's `bind_tools` method is overridden. It typically does not need to be overridden on the test class. Example override: .. code-block:: python @property def has_tool_calling(self) -> bool: return True .. dropdown:: tool_choice_value Value to use for tool choice when used in tests. .. warning:: Deprecated since version 0.3.15: This property will be removed in version 0.3.20. If a model supports ``tool_choice``, it should accept ``tool_choice="any"`` and ``tool_choice=``. If a model does not support forcing tool calling, override the ``has_tool_choice`` property to return ``False``. Example: .. code-block:: python @property def tool_choice_value(self) -> Optional[str]: return "any" .. dropdown:: has_tool_choice Boolean property indicating whether the chat model supports forcing tool calling via a ``tool_choice`` parameter. By default, this is determined by whether the parameter is included in the signature for the corresponding ``bind_tools`` method. If ``True``, the minimum requirement for this feature is that ``tool_choice="any"`` will force a tool call, and ``tool_choice=`` will force a call to a specific tool. Example override: .. code-block:: python @property def has_tool_choice(self) -> bool: return False .. dropdown:: has_structured_output Boolean property indicating whether the chat model supports structured output. By default, this is determined by whether the chat model's ``with_structured_output`` method is overridden. If the base implementation is intended to be used, this method should be overridden. See: https://python.langchain.com/docs/concepts/structured_outputs/ Example: .. code-block:: python @property def has_structured_output(self) -> bool: return True .. dropdown:: structured_output_kwargs Dict property that can be used to specify additional kwargs for ``with_structured_output``. Useful for testing different models. Example: .. code-block:: python @property def structured_output_kwargs(self) -> dict: return {"method": "function_calling"} .. dropdown:: supports_json_mode Boolean property indicating whether the chat model supports JSON mode in ``with_structured_output``. See: https://python.langchain.com/docs/concepts/structured_outputs/#json-mode Example: .. code-block:: python @property def supports_json_mode(self) -> bool: return True .. dropdown:: supports_image_inputs Boolean property indicating whether the chat model supports image inputs. Defaults to ``False``. If set to ``True``, the chat model will be tested using content blocks of the form .. code-block:: python { "type": "image", "source_type": "base64", "data": "", "mime_type": "image/jpeg", # or appropriate mime-type } In addition to OpenAI-style content blocks: .. code-block:: python { "type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}"}, } See https://python.langchain.com/docs/concepts/multimodality/ Example: .. code-block:: python @property def supports_image_inputs(self) -> bool: return True .. dropdown:: supports_image_urls Boolean property indicating whether the chat model supports image inputs from URLs. Defaults to ``False``. If set to ``True``, the chat model will be tested using content blocks of the form .. code-block:: python { "type": "image", "source_type": "url", "url": "https://...", } See https://python.langchain.com/docs/concepts/multimodality/ Example: .. code-block:: python @property def supports_image_urls(self) -> bool: return True .. dropdown:: supports_pdf_inputs Boolean property indicating whether the chat model supports PDF inputs. Defaults to ``False``. If set to ``True``, the chat model will be tested using content blocks of the form .. code-block:: python { "type": "file", "source_type": "base64", "data": "", "mime_type": "application/pdf", } See https://python.langchain.com/docs/concepts/multimodality/ Example: .. code-block:: python @property def supports_pdf_inputs(self) -> bool: return True .. dropdown:: supports_audio_inputs Boolean property indicating whether the chat model supports audio inputs. Defaults to ``False``. If set to ``True``, the chat model will be tested using content blocks of the form .. code-block:: python { "type": "audio", "source_type": "base64", "data": "", "mime_type": "audio/wav", # or appropriate mime-type } See https://python.langchain.com/docs/concepts/multimodality/ Example: .. code-block:: python @property def supports_audio_inputs(self) -> bool: return True .. dropdown:: supports_video_inputs Boolean property indicating whether the chat model supports image inputs. Defaults to ``False``. No current tests are written for this feature. .. dropdown:: returns_usage_metadata Boolean property indicating whether the chat model returns usage metadata on invoke and streaming responses. ``usage_metadata`` is an optional dict attribute on AIMessages that track input and output tokens: https://python.langchain.com/api_reference/core/messages/langchain_core.messages.ai.UsageMetadata.html Example: .. code-block:: python @property def returns_usage_metadata(self) -> bool: return False Models supporting ``usage_metadata`` should also return the name of the underlying model in the ``response_metadata`` of the AIMessage. .. dropdown:: supports_anthropic_inputs Boolean property indicating whether the chat model supports Anthropic-style inputs. These inputs might feature "tool use" and "tool result" content blocks, e.g., .. code-block:: python [ {"type": "text", "text": "Hmm let me think about that"}, { "type": "tool_use", "input": {"fav_color": "green"}, "id": "foo", "name": "color_picker", }, ] If set to ``True``, the chat model will be tested using content blocks of this form. Example: .. code-block:: python @property def supports_anthropic_inputs(self) -> bool: return False .. dropdown:: supports_image_tool_message Boolean property indicating whether the chat model supports ToolMessages that include image content, e.g., .. code-block:: python ToolMessage( content=[ { "type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}"}, }, ], tool_call_id="1", name="random_image", ) (OpenAI Chat Completions format), as well as .. code-block:: python ToolMessage( content=[ { "type": "image", "source_type": "base64", "data": image_data, "mime_type": "image/jpeg", }, ], tool_call_id="1", name="random_image", ) (standard format). If set to ``True``, the chat model will be tested with message sequences that include ToolMessages of this form. Example: .. code-block:: python @property def supports_image_tool_message(self) -> bool: return False .. dropdown:: supported_usage_metadata_details Property controlling what usage metadata details are emitted in both invoke and stream. ``usage_metadata`` is an optional dict attribute on AIMessages that track input and output tokens: https://python.langchain.com/api_reference/core/messages/langchain_core.messages.ai.UsageMetadata.html It includes optional keys ``input_token_details`` and ``output_token_details`` that can track usage details associated with special types of tokens, such as cached, audio, or reasoning. Only needs to be overridden if these details are supplied. .. dropdown:: enable_vcr_tests Property controlling whether to enable select tests that rely on `VCR `_ caching of HTTP calls, such as benchmarking tests. To enable these tests, follow these steps: 1. Override the ``enable_vcr_tests`` property to return ``True``: .. code-block:: python @property def enable_vcr_tests(self) -> bool: return True 2. Configure VCR to exclude sensitive headers and other information from cassettes. .. important:: VCR will by default record authentication headers and other sensitive information in cassettes. Read below for how to configure what information is recorded in cassettes. To add configuration to VCR, add a ``conftest.py`` file to the ``tests/`` directory and implement the ``vcr_config`` fixture there. ``langchain-tests`` excludes the headers ``"authorization"``, ``"x-api-key"``, and ``"api-key"`` from VCR cassettes. To pick up this configuration, you will need to add ``conftest.py`` as shown below. You can also exclude additional headers, override the default exclusions, or apply other customizations to the VCR configuration. See example below: .. code-block:: python :caption: tests/conftest.py import pytest from langchain_tests.conftest import _base_vcr_config as _base_vcr_config _EXTRA_HEADERS = [ # Specify additional headers to redact ("user-agent", "PLACEHOLDER"), ] def remove_response_headers(response: dict) -> dict: # If desired, remove or modify headers in the response. response["headers"] = {} return response @pytest.fixture(scope="session") def vcr_config(_base_vcr_config: dict) -> dict: # noqa: F811 \"\"\"Extend the default configuration from langchain_tests.\"\"\" config = _base_vcr_config.copy() config.setdefault("filter_headers", []).extend(_EXTRA_HEADERS) config["before_record_response"] = remove_response_headers return config .. dropdown:: Compressing cassettes ``langchain-tests`` includes a custom VCR serializer that compresses cassettes using gzip. To use it, register the ``"yaml.gz"`` serializer to your VCR fixture and enable this serializer in the config. See example below: .. code-block:: python :caption: tests/conftest.py import pytest from langchain_tests.conftest import CustomPersister, CustomSerializer from langchain_tests.conftest import _base_vcr_config as _base_vcr_config from vcr import VCR _EXTRA_HEADERS = [ # Specify additional headers to redact ("user-agent", "PLACEHOLDER"), ] def remove_response_headers(response: dict) -> dict: # If desired, remove or modify headers in the response. response["headers"] = {} return response @pytest.fixture(scope="session") def vcr_config(_base_vcr_config: dict) -> dict: # noqa: F811 \"\"\"Extend the default configuration from langchain_tests.\"\"\" config = _base_vcr_config.copy() config.setdefault("filter_headers", []).extend(_EXTRA_HEADERS) config["before_record_response"] = remove_response_headers # New: enable serializer and set file extension config["serializer"] = "yaml.gz" config["path_transformer"] = VCR.ensure_suffix(".yaml.gz") return config def pytest_recording_configure(config: dict, vcr: VCR) -> None: vcr.register_persister(CustomPersister()) vcr.register_serializer("yaml.gz", CustomSerializer()) You can inspect the contents of the compressed cassettes (e.g., to ensure no sensitive information is recorded) using .. code-block:: bash gunzip -k /path/to/tests/cassettes/TestClass_test.yaml.gz or by using the serializer: .. code-block:: python from langchain_tests.conftest import CustomPersister, CustomSerializer cassette_path = "/path/to/tests/cassettes/TestClass_test.yaml.gz" requests, responses = CustomPersister().load_cassette(path, CustomSerializer()) 3. Run tests to generate VCR cassettes. Example: .. code-block:: bash uv run python -m pytest tests/integration_tests/test_chat_models.py::TestMyModel::test_stream_time This will generate a VCR cassette for the test in ``tests/integration_tests/cassettes/``. .. important:: You should inspect the generated cassette to ensure that it does not contain sensitive information. If it does, you can modify the ``vcr_config`` fixture to exclude headers or modify the response before it is recorded. You can then commit the cassette to your repository. Subsequent test runs will use the cassette instead of making HTTP calls. """ # noqa: E501 @property def standard_chat_model_params(self) -> dict: """:private:""" return {} def test_invoke(self, model: BaseChatModel) -> None: """Test to verify that `model.invoke(simple_message)` works. This should pass for all integrations. .. dropdown:: Troubleshooting If this test fails, you should make sure your _generate method does not raise any exceptions, and that it returns a valid :class:`~langchain_core.outputs.chat_result.ChatResult` like so: .. code-block:: python return ChatResult( generations=[ChatGeneration( message=AIMessage(content="Output text") )] ) """ result = model.invoke("Hello") assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 async def test_ainvoke(self, model: BaseChatModel) -> None: """Test to verify that `await model.ainvoke(simple_message)` works. This should pass for all integrations. Passing this test does not indicate a "natively async" implementation, but rather that the model can be used in an async context. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_invoke`. because `ainvoke` has a default implementation that calls `invoke` in an async context. If that test passes but not this one, you should make sure your _agenerate method does not raise any exceptions, and that it returns a valid :class:`~langchain_core.outputs.chat_result.ChatResult` like so: .. code-block:: python return ChatResult( generations=[ChatGeneration( message=AIMessage(content="Output text") )] ) """ result = await model.ainvoke("Hello") assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 def test_stream(self, model: BaseChatModel) -> None: """Test to verify that `model.stream(simple_message)` works. This should pass for all integrations. Passing this test does not indicate a "streaming" implementation, but rather that the model can be used in a streaming context. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_invoke`. because `stream` has a default implementation that calls `invoke` and yields the result as a single chunk. If that test passes but not this one, you should make sure your _stream method does not raise any exceptions, and that it yields valid :class:`~langchain_core.outputs.chat_generation.ChatGenerationChunk` objects like so: .. code-block:: python yield ChatGenerationChunk( message=AIMessageChunk(content="chunk text") ) """ num_tokens = 0 for token in model.stream("Hello"): assert token is not None assert isinstance(token, AIMessageChunk) num_tokens += len(token.content) assert num_tokens > 0 async def test_astream(self, model: BaseChatModel) -> None: """Test to verify that `await model.astream(simple_message)` works. This should pass for all integrations. Passing this test does not indicate a "natively async" or "streaming" implementation, but rather that the model can be used in an async streaming context. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_stream`. and :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_ainvoke`. because `astream` has a default implementation that calls `_stream` in an async context if it is implemented, or `ainvoke` and yields the result as a single chunk if not. If those tests pass but not this one, you should make sure your _astream method does not raise any exceptions, and that it yields valid :class:`~langchain_core.outputs.chat_generation.ChatGenerationChunk` objects like so: .. code-block:: python yield ChatGenerationChunk( message=AIMessageChunk(content="chunk text") ) """ num_tokens = 0 async for token in model.astream("Hello"): assert token is not None assert isinstance(token, AIMessageChunk) num_tokens += len(token.content) assert num_tokens > 0 def test_batch(self, model: BaseChatModel) -> None: """Test to verify that `model.batch([messages])` works. This should pass for all integrations. Tests the model's ability to process multiple prompts in a single batch. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_invoke` because `batch` has a default implementation that calls `invoke` for each message in the batch. If that test passes but not this one, you should make sure your `batch` method does not raise any exceptions, and that it returns a list of valid :class:`~langchain_core.messages.AIMessage` objects. """ batch_results = model.batch(["Hello", "Hey"]) assert batch_results is not None assert isinstance(batch_results, list) assert len(batch_results) == 2 for result in batch_results: assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 async def test_abatch(self, model: BaseChatModel) -> None: """Test to verify that `await model.abatch([messages])` works. This should pass for all integrations. Tests the model's ability to process multiple prompts in a single batch asynchronously. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_batch` and :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_ainvoke` because `abatch` has a default implementation that calls `ainvoke` for each message in the batch. If those tests pass but not this one, you should make sure your `abatch` method does not raise any exceptions, and that it returns a list of valid :class:`~langchain_core.messages.AIMessage` objects. """ batch_results = await model.abatch(["Hello", "Hey"]) assert batch_results is not None assert isinstance(batch_results, list) assert len(batch_results) == 2 for result in batch_results: assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 def test_conversation(self, model: BaseChatModel) -> None: """Test to verify that the model can handle multi-turn conversations. This should pass for all integrations. Tests the model's ability to process a sequence of alternating human and AI messages as context for generating the next response. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_invoke` because this test also uses `model.invoke()`. If that test passes but not this one, you should verify that: 1. Your model correctly processes the message history 2. The model maintains appropriate context from previous messages 3. The response is a valid :class:`~langchain_core.messages.AIMessage` """ messages = [ HumanMessage("hello"), AIMessage("hello"), HumanMessage("how are you"), ] result = model.invoke(messages) assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 def test_double_messages_conversation(self, model: BaseChatModel) -> None: """ Test to verify that the model can handle double-message conversations. This should pass for all integrations. Tests the model's ability to process a sequence of double-system, double-human, and double-ai messages as context for generating the next response. .. dropdown:: Troubleshooting First, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_invoke` because this test also uses `model.invoke()`. Second, debug :meth:`~langchain_tests.integration_tests.chat_models.ChatModelIntegrationTests.test_conversation` because this test is the "basic case" without double messages. If that test passes those but not this one, you should verify that: 1. Your model API can handle double messages, or the integration should merge messages before sending them to the API. 2. The response is a valid :class:`~langchain_core.messages.AIMessage` """ messages = [ SystemMessage("hello"), SystemMessage("hello"), HumanMessage("hello"), HumanMessage("hello"), AIMessage("hello"), AIMessage("hello"), HumanMessage("how are you"), ] result = model.invoke(messages) assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 def test_usage_metadata(self, model: BaseChatModel) -> None: """Test to verify that the model returns correct usage metadata. This test is optional and should be skipped if the model does not return usage metadata (see Configuration below). .. versionchanged:: 0.3.17 Additionally check for the presence of `model_name` in the response metadata, which is needed for usage tracking in callback handlers. .. dropdown:: Configuration By default, this test is run. To disable this feature, set `returns_usage_metadata` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def returns_usage_metadata(self) -> bool: return False This test can also check the format of specific kinds of usage metadata based on the `supported_usage_metadata_details` property. This property should be configured as follows with the types of tokens that the model supports tracking: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supported_usage_metadata_details(self) -> dict: return { "invoke": [ "audio_input", "audio_output", "reasoning_output", "cache_read_input", "cache_creation_input", ], "stream": [ "audio_input", "audio_output", "reasoning_output", "cache_read_input", "cache_creation_input", ], } .. dropdown:: Troubleshooting If this test fails, first verify that your model returns :class:`~langchain_core.messages.ai.UsageMetadata` dicts attached to the returned AIMessage object in `_generate`: .. code-block:: python return ChatResult( generations=[ChatGeneration( message=AIMessage( content="Output text", usage_metadata={ "input_tokens": 350, "output_tokens": 240, "total_tokens": 590, "input_token_details": { "audio": 10, "cache_creation": 200, "cache_read": 100, }, "output_token_details": { "audio": 10, "reasoning": 200, } } ) )] ) Check also that the response includes a ``"model_name"`` key in its ``usage_metadata``. """ if not self.returns_usage_metadata: pytest.skip("Not implemented.") result = model.invoke("Hello") assert result is not None assert isinstance(result, AIMessage) assert result.usage_metadata is not None assert isinstance(result.usage_metadata["input_tokens"], int) assert isinstance(result.usage_metadata["output_tokens"], int) assert isinstance(result.usage_metadata["total_tokens"], int) # Check model_name is in response_metadata # Needed for langchain_core.callbacks.usage model_name = result.response_metadata.get("model_name") assert isinstance(model_name, str) assert model_name if "audio_input" in self.supported_usage_metadata_details["invoke"]: msg = self.invoke_with_audio_input() assert msg.usage_metadata is not None assert msg.usage_metadata["input_token_details"] is not None assert isinstance(msg.usage_metadata["input_token_details"]["audio"], int) assert msg.usage_metadata["input_tokens"] >= sum( (v or 0) # type: ignore[misc] for v in msg.usage_metadata["input_token_details"].values() ) if "audio_output" in self.supported_usage_metadata_details["invoke"]: msg = self.invoke_with_audio_output() assert msg.usage_metadata is not None assert msg.usage_metadata["output_token_details"] is not None assert isinstance(msg.usage_metadata["output_token_details"]["audio"], int) assert int(msg.usage_metadata["output_tokens"]) >= sum( (v or 0) # type: ignore[misc] for v in msg.usage_metadata["output_token_details"].values() ) if "reasoning_output" in self.supported_usage_metadata_details["invoke"]: msg = self.invoke_with_reasoning_output() assert msg.usage_metadata is not None assert msg.usage_metadata["output_token_details"] is not None assert isinstance( msg.usage_metadata["output_token_details"]["reasoning"], int, ) assert msg.usage_metadata["output_tokens"] >= sum( (v or 0) # type: ignore[misc] for v in msg.usage_metadata["output_token_details"].values() ) if "cache_read_input" in self.supported_usage_metadata_details["invoke"]: msg = self.invoke_with_cache_read_input() assert msg.usage_metadata is not None assert msg.usage_metadata["input_token_details"] is not None assert isinstance( msg.usage_metadata["input_token_details"]["cache_read"], int, ) assert msg.usage_metadata["input_tokens"] >= sum( (v or 0) # type: ignore[misc] for v in msg.usage_metadata["input_token_details"].values() ) if "cache_creation_input" in self.supported_usage_metadata_details["invoke"]: msg = self.invoke_with_cache_creation_input() assert msg.usage_metadata is not None assert msg.usage_metadata["input_token_details"] is not None assert isinstance( msg.usage_metadata["input_token_details"]["cache_creation"], int, ) assert msg.usage_metadata["input_tokens"] >= sum( (v or 0) # type: ignore[misc] for v in msg.usage_metadata["input_token_details"].values() ) def test_usage_metadata_streaming(self, model: BaseChatModel) -> None: """ Test to verify that the model returns correct usage metadata in streaming mode. .. versionchanged:: 0.3.17 Additionally check for the presence of `model_name` in the response metadata, which is needed for usage tracking in callback handlers. .. dropdown:: Configuration By default, this test is run. To disable this feature, set `returns_usage_metadata` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def returns_usage_metadata(self) -> bool: return False This test can also check the format of specific kinds of usage metadata based on the `supported_usage_metadata_details` property. This property should be configured as follows with the types of tokens that the model supports tracking: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supported_usage_metadata_details(self) -> dict: return { "invoke": [ "audio_input", "audio_output", "reasoning_output", "cache_read_input", "cache_creation_input", ], "stream": [ "audio_input", "audio_output", "reasoning_output", "cache_read_input", "cache_creation_input", ], } .. dropdown:: Troubleshooting If this test fails, first verify that your model yields :class:`~langchain_core.messages.ai.UsageMetadata` dicts attached to the returned AIMessage object in `_stream` that sum up to the total usage metadata. Note that `input_tokens` should only be included on one of the chunks (typically the first or the last chunk), and the rest should have 0 or None to avoid counting input tokens multiple times. `output_tokens` typically count the number of tokens in each chunk, not the sum. This test will pass as long as the sum of `output_tokens` across all chunks is not 0. .. code-block:: python yield ChatResult( generations=[ChatGeneration( message=AIMessage( content="Output text", usage_metadata={ "input_tokens": ( num_input_tokens if is_first_chunk else 0 ), "output_tokens": 11, "total_tokens": ( 11+num_input_tokens if is_first_chunk else 11 ), "input_token_details": { "audio": 10, "cache_creation": 200, "cache_read": 100, }, "output_token_details": { "audio": 10, "reasoning": 200, } } ) )] ) Check also that the aggregated response includes a ``"model_name"`` key in its ``usage_metadata``. """ if not self.returns_usage_metadata: pytest.skip("Not implemented.") full: Optional[AIMessageChunk] = None for chunk in model.stream("Write me 2 haikus. Only include the haikus."): assert isinstance(chunk, AIMessageChunk) # only one chunk is allowed to set usage_metadata.input_tokens # if multiple do, it's likely a bug that will result in overcounting # input tokens if full and full.usage_metadata and full.usage_metadata["input_tokens"]: assert ( not chunk.usage_metadata or not chunk.usage_metadata["input_tokens"] ), ( "Only one chunk should set input_tokens," " the rest should be 0 or None" ) full = chunk if full is None else cast(AIMessageChunk, full + chunk) assert isinstance(full, AIMessageChunk) assert full.usage_metadata is not None assert isinstance(full.usage_metadata["input_tokens"], int) assert isinstance(full.usage_metadata["output_tokens"], int) assert isinstance(full.usage_metadata["total_tokens"], int) # Check model_name is in response_metadata # Needed for langchain_core.callbacks.usage model_name = full.response_metadata.get("model_name") assert isinstance(model_name, str) assert model_name if "audio_input" in self.supported_usage_metadata_details["stream"]: msg = self.invoke_with_audio_input(stream=True) assert isinstance(msg.usage_metadata["input_token_details"]["audio"], int) # type: ignore[index] if "audio_output" in self.supported_usage_metadata_details["stream"]: msg = self.invoke_with_audio_output(stream=True) assert isinstance(msg.usage_metadata["output_token_details"]["audio"], int) # type: ignore[index] if "reasoning_output" in self.supported_usage_metadata_details["stream"]: msg = self.invoke_with_reasoning_output(stream=True) assert isinstance( msg.usage_metadata["output_token_details"]["reasoning"], # type: ignore[index] int, ) if "cache_read_input" in self.supported_usage_metadata_details["stream"]: msg = self.invoke_with_cache_read_input(stream=True) assert isinstance( msg.usage_metadata["input_token_details"]["cache_read"], # type: ignore[index] int, ) if "cache_creation_input" in self.supported_usage_metadata_details["stream"]: msg = self.invoke_with_cache_creation_input(stream=True) assert isinstance( msg.usage_metadata["input_token_details"]["cache_creation"], # type: ignore[index] int, ) def test_stop_sequence(self, model: BaseChatModel) -> None: """Test that model does not fail when invoked with the ``stop`` parameter, which is a standard parameter for stopping generation at a certain token. More on standard parameters here: https://python.langchain.com/docs/concepts/chat_models/#standard-parameters This should pass for all integrations. .. dropdown:: Troubleshooting If this test fails, check that the function signature for ``_generate`` (as well as ``_stream`` and async variants) accepts the ``stop`` parameter: .. code-block:: python def _generate( self, messages: List[BaseMessage], stop: Optional[List[str]] = None, run_manager: Optional[CallbackManagerForLLMRun] = None, **kwargs: Any, ) -> ChatResult: """ # noqa: E501 result = model.invoke("hi", stop=["you"]) assert isinstance(result, AIMessage) custom_model = self.chat_model_class( **{ **self.chat_model_params, "stop": ["you"], } ) result = custom_model.invoke("hi") assert isinstance(result, AIMessage) def test_tool_calling(self, model: BaseChatModel) -> None: """Test that the model generates tool calls. This test is skipped if the ``has_tool_calling`` property on the test class is set to False. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that ``bind_tools`` is implemented to correctly translate LangChain tool objects into the appropriate schema for your chat model. This test may fail if the chat model does not support a ``tool_choice`` parameter. This parameter can be used to force a tool call. If ``tool_choice`` is not supported and the model consistently fails this test, you can ``xfail`` the test: .. code-block:: python @pytest.mark.xfail(reason=("Does not support tool_choice.")) def test_tool_calling(self, model: BaseChatModel) -> None: super().test_tool_calling(model) Otherwise, in the case that only one tool is bound, ensure that ``tool_choice`` supports the string ``"any"`` to force calling that tool. """ if not self.has_tool_calling: pytest.skip("Test requires tool calling.") if not self.has_tool_choice: tool_choice_value = None else: tool_choice_value = "any" # Emit warning if tool_choice_value property is overridden if inspect.getattr_static( self, "tool_choice_value" ) is not inspect.getattr_static(ChatModelIntegrationTests, "tool_choice_value"): warn_deprecated( "0.3.15", message=( "`tool_choice_value` will be removed in version 0.3.20. If a " "model supports `tool_choice`, it should accept `tool_choice='any' " "and `tool_choice=`. If the model does not " "support `tool_choice`, override the `supports_tool_choice` " "property to return `False`." ), removal="0.3.20", ) model_with_tools = model.bind_tools( [magic_function], tool_choice=tool_choice_value ) # Test invoke query = "What is the value of magic_function(3)? Use the tool." result = model_with_tools.invoke(query) _validate_tool_call_message(result) # Test stream full: Optional[BaseMessageChunk] = None for chunk in model_with_tools.stream(query): full = chunk if full is None else full + chunk # type: ignore assert isinstance(full, AIMessage) _validate_tool_call_message(full) def test_tool_choice(self, model: BaseChatModel) -> None: """Test that the model can force tool calling via the ``tool_choice`` parameter. This test is skipped if the ``has_tool_choice`` property on the test class is set to False. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_choice`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_choice(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check whether the ``test_tool_calling`` test is passing. If it is not, refer to the troubleshooting steps in that test first. If ``test_tool_calling`` is passing, check that the underlying model supports forced tool calling. If it does, ``bind_tools`` should accept a ``tool_choice`` parameter that can be used to force a tool call. It should accept (1) the string ``"any"`` to force calling the bound tool, and (2) the string name of the tool to force calling that tool. """ if not self.has_tool_choice or not self.has_tool_calling: pytest.skip("Test requires tool choice.") @tool def get_weather(location: str) -> str: """Get weather at a location.""" return "It's sunny." for tool_choice in ["any", "magic_function"]: model_with_tools = model.bind_tools( [magic_function, get_weather], tool_choice=tool_choice ) result = model_with_tools.invoke("Hello!") assert isinstance(result, AIMessage) assert result.tool_calls if tool_choice == "magic_function": assert result.tool_calls[0]["name"] == "magic_function" async def test_tool_calling_async(self, model: BaseChatModel) -> None: """Test that the model generates tool calls. This test is skipped if the ``has_tool_calling`` property on the test class is set to False. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that ``bind_tools`` is implemented to correctly translate LangChain tool objects into the appropriate schema for your chat model. This test may fail if the chat model does not support a ``tool_choice`` parameter. This parameter can be used to force a tool call. If ``tool_choice`` is not supported and the model consistently fails this test, you can ``xfail`` the test: .. code-block:: python @pytest.mark.xfail(reason=("Does not support tool_choice.")) async def test_tool_calling_async(self, model: BaseChatModel) -> None: await super().test_tool_calling_async(model) Otherwise, in the case that only one tool is bound, ensure that ``tool_choice`` supports the string ``"any"`` to force calling that tool. """ if not self.has_tool_calling: pytest.skip("Test requires tool calling.") if not self.has_tool_choice: tool_choice_value = None else: tool_choice_value = "any" model_with_tools = model.bind_tools( [magic_function], tool_choice=tool_choice_value ) # Test ainvoke query = "What is the value of magic_function(3)? Use the tool." result = await model_with_tools.ainvoke(query) _validate_tool_call_message(result) # Test astream full: Optional[BaseMessageChunk] = None async for chunk in model_with_tools.astream(query): full = chunk if full is None else full + chunk # type: ignore assert isinstance(full, AIMessage) _validate_tool_call_message(full) def test_tool_calling_with_no_arguments(self, model: BaseChatModel) -> None: """Test that the model generates tool calls for tools with no arguments. This test is skipped if the ``has_tool_calling`` property on the test class is set to False. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that ``bind_tools`` is implemented to correctly translate LangChain tool objects into the appropriate schema for your chat model. It should correctly handle the case where a tool has no arguments. This test may fail if the chat model does not support a ``tool_choice`` parameter. This parameter can be used to force a tool call. It may also fail if a provider does not support this form of tool. In these cases, you can ``xfail`` the test: .. code-block:: python @pytest.mark.xfail(reason=("Does not support tool_choice.")) def test_tool_calling_with_no_arguments(self, model: BaseChatModel) -> None: super().test_tool_calling_with_no_arguments(model) Otherwise, in the case that only one tool is bound, ensure that ``tool_choice`` supports the string ``"any"`` to force calling that tool. """ # noqa: E501 if not self.has_tool_calling: pytest.skip("Test requires tool calling.") if not self.has_tool_choice: tool_choice_value = None else: tool_choice_value = "any" model_with_tools = model.bind_tools( [magic_function_no_args], tool_choice=tool_choice_value ) query = "What is the value of magic_function_no_args()? Use the tool." result = model_with_tools.invoke(query) _validate_tool_call_message_no_args(result) full: Optional[BaseMessageChunk] = None for chunk in model_with_tools.stream(query): full = chunk if full is None else full + chunk # type: ignore assert isinstance(full, AIMessage) _validate_tool_call_message_no_args(full) def test_bind_runnables_as_tools(self, model: BaseChatModel) -> None: """Test that the model generates tool calls for tools that are derived from LangChain runnables. This test is skipped if the ``has_tool_calling`` property on the test class is set to False. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that ``bind_tools`` is implemented to correctly translate LangChain tool objects into the appropriate schema for your chat model. This test may fail if the chat model does not support a ``tool_choice`` parameter. This parameter can be used to force a tool call. If ``tool_choice`` is not supported and the model consistently fails this test, you can ``xfail`` the test: .. code-block:: python @pytest.mark.xfail(reason=("Does not support tool_choice.")) def test_bind_runnables_as_tools(self, model: BaseChatModel) -> None: super().test_bind_runnables_as_tools(model) Otherwise, ensure that the ``tool_choice_value`` property is correctly specified on the test class. """ if not self.has_tool_calling: pytest.skip("Test requires tool calling.") prompt = ChatPromptTemplate.from_messages( [("human", "Hello. Please respond in the style of {answer_style}.")] ) llm = GenericFakeChatModel(messages=iter(["hello matey"])) chain = prompt | llm | StrOutputParser() tool_ = chain.as_tool( name="greeting_generator", description="Generate a greeting in a particular style of speaking.", ) if self.has_tool_choice: tool_choice: Optional[str] = "any" else: tool_choice = None model_with_tools = model.bind_tools([tool_], tool_choice=tool_choice) query = "Using the tool, generate a Pirate greeting." result = model_with_tools.invoke(query) assert isinstance(result, AIMessage) assert result.tool_calls tool_call = result.tool_calls[0] assert tool_call["args"].get("answer_style") assert tool_call["type"] == "tool_call" @pytest.mark.parametrize("schema_type", ["pydantic", "typeddict", "json_schema"]) def test_structured_output(self, model: BaseChatModel, schema_type: str) -> None: """Test to verify structured output is generated both on invoke and stream. This test is optional and should be skipped if the model does not support structured output (see Configuration below). .. dropdown:: Configuration To disable structured output tests, set ``has_structured_output`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_structured_output(self) -> bool: return False By default, ``has_structured_output`` is True if a model overrides the ``with_structured_output`` or ``bind_tools`` methods. .. dropdown:: Troubleshooting If this test fails, ensure that the model's ``bind_tools`` method properly handles both JSON Schema and Pydantic V2 models. ``langchain_core`` implements a utility function that will accommodate most formats: https://python.langchain.com/api_reference/core/utils/langchain_core.utils.function_calling.convert_to_openai_tool.html See example implementation of ``with_structured_output`` here: https://python.langchain.com/api_reference/_modules/langchain_openai/chat_models/base.html#BaseChatOpenAI.with_structured_output """ # noqa: E501 if not self.has_structured_output: pytest.skip("Test requires structured output.") schema, validation_function = _get_joke_class(schema_type) # type: ignore[arg-type] chat = model.with_structured_output(schema, **self.structured_output_kwargs) mock_callback = MagicMock() mock_callback.on_chat_model_start = MagicMock() invoke_callback = _TestCallbackHandler() result = chat.invoke( "Tell me a joke about cats.", config={"callbacks": [invoke_callback]} ) validation_function(result) assert len(invoke_callback.options) == 1, ( "Expected on_chat_model_start to be called once" ) assert isinstance(invoke_callback.options[0], dict) assert isinstance( invoke_callback.options[0]["ls_structured_output_format"]["schema"], dict ) assert invoke_callback.options[0]["ls_structured_output_format"][ "schema" ] == convert_to_json_schema(schema) stream_callback = _TestCallbackHandler() for chunk in chat.stream( "Tell me a joke about cats.", config={"callbacks": [stream_callback]} ): validation_function(chunk) assert chunk assert len(stream_callback.options) == 1, ( "Expected on_chat_model_start to be called once" ) assert isinstance(stream_callback.options[0], dict) assert isinstance( stream_callback.options[0]["ls_structured_output_format"]["schema"], dict ) assert stream_callback.options[0]["ls_structured_output_format"][ "schema" ] == convert_to_json_schema(schema) @pytest.mark.parametrize("schema_type", ["pydantic", "typeddict", "json_schema"]) async def test_structured_output_async( self, model: BaseChatModel, schema_type: str ) -> None: """Test to verify structured output is generated both on invoke and stream. This test is optional and should be skipped if the model does not support structured output (see Configuration below). .. dropdown:: Configuration To disable structured output tests, set ``has_structured_output`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_structured_output(self) -> bool: return False By default, ``has_structured_output`` is True if a model overrides the ``with_structured_output`` or ``bind_tools`` methods. .. dropdown:: Troubleshooting If this test fails, ensure that the model's ``bind_tools`` method properly handles both JSON Schema and Pydantic V2 models. ``langchain_core`` implements a utility function that will accommodate most formats: https://python.langchain.com/api_reference/core/utils/langchain_core.utils.function_calling.convert_to_openai_tool.html See example implementation of ``with_structured_output`` here: https://python.langchain.com/api_reference/_modules/langchain_openai/chat_models/base.html#BaseChatOpenAI.with_structured_output """ # noqa: E501 if not self.has_structured_output: pytest.skip("Test requires structured output.") schema, validation_function = _get_joke_class(schema_type) # type: ignore[arg-type] chat = model.with_structured_output(schema, **self.structured_output_kwargs) ainvoke_callback = _TestCallbackHandler() result = await chat.ainvoke( "Tell me a joke about cats.", config={"callbacks": [ainvoke_callback]} ) validation_function(result) assert len(ainvoke_callback.options) == 1, ( "Expected on_chat_model_start to be called once" ) assert isinstance(ainvoke_callback.options[0], dict) assert isinstance( ainvoke_callback.options[0]["ls_structured_output_format"]["schema"], dict ) assert ainvoke_callback.options[0]["ls_structured_output_format"][ "schema" ] == convert_to_json_schema(schema) astream_callback = _TestCallbackHandler() async for chunk in chat.astream( "Tell me a joke about cats.", config={"callbacks": [astream_callback]} ): validation_function(chunk) assert chunk assert len(astream_callback.options) == 1, ( "Expected on_chat_model_start to be called once" ) assert isinstance(astream_callback.options[0], dict) assert isinstance( astream_callback.options[0]["ls_structured_output_format"]["schema"], dict ) assert astream_callback.options[0]["ls_structured_output_format"][ "schema" ] == convert_to_json_schema(schema) @pytest.mark.skipif(PYDANTIC_MAJOR_VERSION != 2, reason="Test requires pydantic 2.") def test_structured_output_pydantic_2_v1(self, model: BaseChatModel) -> None: """Test to verify we can generate structured output using pydantic.v1.BaseModel. pydantic.v1.BaseModel is available in the pydantic 2 package. This test is optional and should be skipped if the model does not support structured output (see Configuration below). .. dropdown:: Configuration To disable structured output tests, set ``has_structured_output`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_structured_output(self) -> bool: return False By default, ``has_structured_output`` is True if a model overrides the ``with_structured_output`` or ``bind_tools`` methods. .. dropdown:: Troubleshooting If this test fails, ensure that the model's ``bind_tools`` method properly handles both JSON Schema and Pydantic V1 models. ``langchain_core`` implements a utility function that will accommodate most formats: https://python.langchain.com/api_reference/core/utils/langchain_core.utils.function_calling.convert_to_openai_tool.html See example implementation of ``with_structured_output`` here: https://python.langchain.com/api_reference/_modules/langchain_openai/chat_models/base.html#BaseChatOpenAI.with_structured_output """ if not self.has_structured_output: pytest.skip("Test requires structured output.") class Joke(BaseModelV1): # Uses langchain_core.pydantic_v1.BaseModel """Joke to tell user.""" setup: str = FieldV1(description="question to set up a joke") punchline: str = FieldV1(description="answer to resolve the joke") # Pydantic class chat = model.with_structured_output(Joke, **self.structured_output_kwargs) result = chat.invoke("Tell me a joke about cats.") assert isinstance(result, Joke) for chunk in chat.stream("Tell me a joke about cats."): assert isinstance(chunk, Joke) # Schema chat = model.with_structured_output( Joke.schema(), **self.structured_output_kwargs ) result = chat.invoke("Tell me a joke about cats.") assert isinstance(result, dict) assert set(result.keys()) == {"setup", "punchline"} for chunk in chat.stream("Tell me a joke about cats."): assert isinstance(chunk, dict) assert isinstance(chunk, dict) # for mypy assert set(chunk.keys()) == {"setup", "punchline"} def test_structured_output_optional_param(self, model: BaseChatModel) -> None: """Test to verify we can generate structured output that includes optional parameters. This test is optional and should be skipped if the model does not support structured output (see Configuration below). .. dropdown:: Configuration To disable structured output tests, set ``has_structured_output`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_structured_output(self) -> bool: return False By default, ``has_structured_output`` is True if a model overrides the ``with_structured_output`` or ``bind_tools`` methods. .. dropdown:: Troubleshooting If this test fails, ensure that the model's ``bind_tools`` method properly handles Pydantic V2 models with optional parameters. ``langchain_core`` implements a utility function that will accommodate most formats: https://python.langchain.com/api_reference/core/utils/langchain_core.utils.function_calling.convert_to_openai_tool.html See example implementation of ``with_structured_output`` here: https://python.langchain.com/api_reference/_modules/langchain_openai/chat_models/base.html#BaseChatOpenAI.with_structured_output """ if not self.has_structured_output: pytest.skip("Test requires structured output.") # Pydantic class Joke(BaseModel): """Joke to tell user.""" setup: str = Field(description="question to set up a joke") punchline: Optional[str] = Field( default=None, description="answer to resolve the joke" ) chat = model.with_structured_output(Joke, **self.structured_output_kwargs) setup_result = chat.invoke( "Give me the setup to a joke about cats, no punchline." ) assert isinstance(setup_result, Joke) joke_result = chat.invoke("Give me a joke about cats, include the punchline.") assert isinstance(joke_result, Joke) # Schema chat = model.with_structured_output( Joke.model_json_schema(), **self.structured_output_kwargs ) result = chat.invoke("Tell me a joke about cats.") assert isinstance(result, dict) # TypedDict class JokeDict(TypedDict): """Joke to tell user.""" setup: Annotated[str, ..., "question to set up a joke"] punchline: Annotated[Optional[str], None, "answer to resolve the joke"] chat = model.with_structured_output(JokeDict, **self.structured_output_kwargs) result = chat.invoke("Tell me a joke about cats.") assert isinstance(result, dict) def test_json_mode(self, model: BaseChatModel) -> None: """Test structured output via `JSON mode. `_ This test is optional and should be skipped if the model does not support the JSON mode feature (see Configuration below). .. dropdown:: Configuration To disable this test, set ``supports_json_mode`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supports_json_mode(self) -> bool: return False .. dropdown:: Troubleshooting See example implementation of ``with_structured_output`` here: https://python.langchain.com/api_reference/_modules/langchain_openai/chat_models/base.html#BaseChatOpenAI.with_structured_output """ # noqa: E501 if not self.supports_json_mode: pytest.skip("Test requires json mode support.") from pydantic import BaseModel as BaseModelProper from pydantic import Field as FieldProper class Joke(BaseModelProper): """Joke to tell user.""" setup: str = FieldProper(description="question to set up a joke") punchline: str = FieldProper(description="answer to resolve the joke") # Pydantic class # Type ignoring since the interface only officially supports pydantic 1 # or pydantic.v1.BaseModel but not pydantic.BaseModel from pydantic 2. # We'll need to do a pass updating the type signatures. chat = model.with_structured_output(Joke, method="json_mode") # type: ignore[arg-type] msg = ( "Tell me a joke about cats. Return the result as a JSON with 'setup' and " "'punchline' keys. Return nothing other than JSON." ) result = chat.invoke(msg) assert isinstance(result, Joke) for chunk in chat.stream(msg): assert isinstance(chunk, Joke) # Schema chat = model.with_structured_output( Joke.model_json_schema(), method="json_mode" ) result = chat.invoke(msg) assert isinstance(result, dict) assert set(result.keys()) == {"setup", "punchline"} for chunk in chat.stream(msg): assert isinstance(chunk, dict) assert isinstance(chunk, dict) # for mypy assert set(chunk.keys()) == {"setup", "punchline"} def test_tool_message_histories_string_content( self, model: BaseChatModel, my_adder_tool: BaseTool ) -> None: """Test that message histories are compatible with string tool contents (e.g. OpenAI format). If a model passes this test, it should be compatible with messages generated from providers following OpenAI format. This test should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that: 1. The model can correctly handle message histories that include AIMessage objects with ``""`` content. 2. The ``tool_calls`` attribute on AIMessage objects is correctly handled and passed to the model in an appropriate format. 3. The model can correctly handle ToolMessage objects with string content and arbitrary string values for ``tool_call_id``. You can ``xfail`` the test if tool calling is implemented but this format is not supported. .. code-block:: python @pytest.mark.xfail(reason=("Not implemented.")) def test_tool_message_histories_string_content(self, *args: Any) -> None: super().test_tool_message_histories_string_content(*args) """ # noqa: E501 if not self.has_tool_calling: pytest.skip("Test requires tool calling.") model_with_tools = model.bind_tools([my_adder_tool]) function_name = "my_adder_tool" function_args = {"a": "1", "b": "2"} messages_string_content = [ HumanMessage("What is 1 + 2"), # string content (e.g. OpenAI) AIMessage( "", tool_calls=[ { "name": function_name, "args": function_args, "id": "abc123", "type": "tool_call", }, ], ), ToolMessage( json.dumps({"result": 3}), name=function_name, tool_call_id="abc123", ), ] result_string_content = model_with_tools.invoke(messages_string_content) assert isinstance(result_string_content, AIMessage) def test_tool_message_histories_list_content( self, model: BaseChatModel, my_adder_tool: BaseTool, ) -> None: """Test that message histories are compatible with list tool contents (e.g. Anthropic format). These message histories will include AIMessage objects with "tool use" and content blocks, e.g., .. code-block:: python [ {"type": "text", "text": "Hmm let me think about that"}, { "type": "tool_use", "input": {"fav_color": "green"}, "id": "foo", "name": "color_picker", }, ] This test should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that: 1. The model can correctly handle message histories that include AIMessage objects with list content. 2. The ``tool_calls`` attribute on AIMessage objects is correctly handled and passed to the model in an appropriate format. 3. The model can correctly handle ToolMessage objects with string content and arbitrary string values for ``tool_call_id``. You can ``xfail`` the test if tool calling is implemented but this format is not supported. .. code-block:: python @pytest.mark.xfail(reason=("Not implemented.")) def test_tool_message_histories_list_content(self, *args: Any) -> None: super().test_tool_message_histories_list_content(*args) """ # noqa: E501 if not self.has_tool_calling: pytest.skip("Test requires tool calling.") model_with_tools = model.bind_tools([my_adder_tool]) function_name = "my_adder_tool" function_args = {"a": 1, "b": 2} messages_list_content = [ HumanMessage("What is 1 + 2"), # List content (e.g., Anthropic) AIMessage( [ {"type": "text", "text": "some text"}, { "type": "tool_use", "id": "abc123", "name": function_name, "input": function_args, }, ], tool_calls=[ { "name": function_name, "args": function_args, "id": "abc123", "type": "tool_call", }, ], ), ToolMessage( json.dumps({"result": 3}), name=function_name, tool_call_id="abc123", ), ] result_list_content = model_with_tools.invoke(messages_list_content) assert isinstance(result_list_content, AIMessage) def test_structured_few_shot_examples( self, model: BaseChatModel, my_adder_tool: BaseTool ) -> None: """Test that the model can process few-shot examples with tool calls. These are represented as a sequence of messages of the following form: - ``HumanMessage`` with string content; - ``AIMessage`` with the ``tool_calls`` attribute populated; - ``ToolMessage`` with string content; - ``AIMessage`` with string content (an answer); - ``HuamnMessage`` with string content (a follow-up question). This test should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting This test uses a utility function in ``langchain_core`` to generate a sequence of messages representing "few-shot" examples: https://python.langchain.com/api_reference/core/utils/langchain_core.utils.function_calling.tool_example_to_messages.html If this test fails, check that the model can correctly handle this sequence of messages. You can ``xfail`` the test if tool calling is implemented but this format is not supported. .. code-block:: python @pytest.mark.xfail(reason=("Not implemented.")) def test_structured_few_shot_examples(self, *args: Any) -> None: super().test_structured_few_shot_examples(*args) """ # noqa: E501 if not self.has_tool_calling: pytest.skip("Test requires tool calling.") model_with_tools = model.bind_tools([my_adder_tool], tool_choice="any") function_result = json.dumps({"result": 3}) tool_schema = my_adder_tool.args_schema assert isinstance(tool_schema, type) and issubclass(tool_schema, BaseModel) few_shot_messages = tool_example_to_messages( "What is 1 + 2", [tool_schema(a=1, b=2)], tool_outputs=[function_result], ai_response=function_result, ) messages = few_shot_messages + [HumanMessage("What is 3 + 4")] result = model_with_tools.invoke(messages) assert isinstance(result, AIMessage) def test_pdf_inputs(self, model: BaseChatModel) -> None: """Test that the model can process PDF inputs. This test should be skipped (see Configuration below) if the model does not support PDF inputs. These will take the form: .. code-block:: python { "type": "image", "source_type": "base64", "data": "", "mime_type": "application/pdf", } See https://python.langchain.com/docs/concepts/multimodality/ .. dropdown:: Configuration To disable this test, set ``supports_pdf_inputs`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supports_pdf_inputs(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that the model can correctly handle messages with pdf content blocks, including base64-encoded files. Otherwise, set the ``supports_pdf_inputs`` property to False. """ if not self.supports_pdf_inputs: pytest.skip("Model does not support PDF inputs.") url = "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf" pdf_data = base64.b64encode(httpx.get(url).content).decode("utf-8") message = HumanMessage( [ { "type": "text", "text": "Summarize this document:", }, { "type": "file", "source_type": "base64", "mime_type": "application/pdf", "data": pdf_data, }, ] ) _ = model.invoke([message]) # Test OpenAI Chat Completions format message = HumanMessage( [ { "type": "text", "text": "Summarize this document:", }, { "type": "file", "file": { "filename": "test file.pdf", "file_data": f"data:application/pdf;base64,{pdf_data}", }, }, ] ) _ = model.invoke([message]) def test_audio_inputs(self, model: BaseChatModel) -> None: """Test that the model can process audio inputs. This test should be skipped (see Configuration below) if the model does not support audio inputs. These will take the form: .. code-block:: python { "type": "audio", "source_type": "base64", "data": "", "mime_type": "audio/wav", # or appropriate mime-type } See https://python.langchain.com/docs/concepts/multimodality/ .. dropdown:: Configuration To disable this test, set ``supports_audio_inputs`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supports_audio_inputs(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that the model can correctly handle messages with audio content blocks, specifically base64-encoded files. Otherwise, set the ``supports_audio_inputs`` property to False. """ if not self.supports_audio_inputs: pytest.skip("Model does not support audio inputs.") url = "https://upload.wikimedia.org/wikipedia/commons/3/3d/Alcal%C3%A1_de_Henares_%28RPS_13-04-2024%29_canto_de_ruise%C3%B1or_%28Luscinia_megarhynchos%29_en_el_Soto_del_Henares.wav" audio_data = base64.b64encode(httpx.get(url).content).decode("utf-8") message = HumanMessage( [ { "type": "text", "text": "Describe this audio:", }, { "type": "audio", "source_type": "base64", "mime_type": "audio/wav", "data": audio_data, }, ] ) _ = model.invoke([message]) # Test OpenAI Chat Completions format message = HumanMessage( [ { "type": "text", "text": "Describe this audio:", }, { "type": "input_audio", "input_audio": {"data": audio_data, "format": "wav"}, }, ] ) _ = model.invoke([message]) def test_image_inputs(self, model: BaseChatModel) -> None: """Test that the model can process image inputs. This test should be skipped (see Configuration below) if the model does not support image inputs. These will take the form: .. code-block:: python { "type": "image", "source_type": "base64", "data": "", "mime_type": "image/jpeg", # or appropriate mime-type } For backward-compatibility, we must also support OpenAI-style image content blocks: .. code-block:: python [ {"type": "text", "text": "describe the weather in this image"}, { "type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}"}, }, ] See https://python.langchain.com/docs/concepts/multimodality/ If the property ``supports_image_urls`` is set to True, the test will also check that we can process content blocks of the form: .. code-block:: python { "type": "image", "source_type": "url", "url": "", } .. dropdown:: Configuration To disable this test, set ``supports_image_inputs`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supports_image_inputs(self) -> bool: return False # Can also explicitly disable testing image URLs: @property def supports_image_urls(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that the model can correctly handle messages with image content blocks, including base64-encoded images. Otherwise, set the ``supports_image_inputs`` property to False. """ if not self.supports_image_inputs: pytest.skip("Model does not support image message.") image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" image_data = base64.b64encode(httpx.get(image_url).content).decode("utf-8") # OpenAI format, base64 data message = HumanMessage( content=[ {"type": "text", "text": "describe the weather in this image"}, { "type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}"}, }, ], ) _ = model.invoke([message]) # Standard format, base64 data message = HumanMessage( content=[ {"type": "text", "text": "describe the weather in this image"}, { "type": "image", "source_type": "base64", "mime_type": "image/jpeg", "data": image_data, }, ], ) _ = model.invoke([message]) # Standard format, URL if self.supports_image_urls: message = HumanMessage( content=[ {"type": "text", "text": "describe the weather in this image"}, { "type": "image", "source_type": "url", "url": image_url, }, ], ) _ = model.invoke([message]) def test_image_tool_message(self, model: BaseChatModel) -> None: """Test that the model can process ToolMessages with image inputs. This test should be skipped if the model does not support messages of the form: .. code-block:: python ToolMessage( content=[ { "type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}"}, }, ], tool_call_id="1", name="random_image", ) containing image content blocks in OpenAI Chat Completions format, in addition to messages of the form: .. code-block:: python ToolMessage( content=[ { "type": "image", "source_type": "base64", "data": image_data, "mime_type": "image/jpeg", }, ], tool_call_id="1", name="random_image", ) containing image content blocks in standard format. This test can be skipped by setting the ``supports_image_tool_message`` property to False (see Configuration below). .. dropdown:: Configuration To disable this test, set ``supports_image_tool_message`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supports_image_tool_message(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that the model can correctly handle messages with image content blocks in ToolMessages, including base64-encoded images. Otherwise, set the ``supports_image_tool_message`` property to False. """ if not self.supports_image_tool_message: pytest.skip("Model does not support image tool message.") image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" image_data = base64.b64encode(httpx.get(image_url).content).decode("utf-8") # Support both OpenAI and standard formats oai_format_message = ToolMessage( content=[ { "type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}"}, }, ], tool_call_id="1", name="random_image", ) standard_format_message = ToolMessage( content=[ { "type": "image", "source_type": "base64", "data": image_data, "mime_type": "image/jpeg", }, ], tool_call_id="1", name="random_image", ) for tool_message in [oai_format_message, standard_format_message]: messages = [ HumanMessage( "get a random image using the tool and describe the weather" ), AIMessage( [], tool_calls=[ { "type": "tool_call", "id": "1", "name": "random_image", "args": {}, } ], ), tool_message, ] def random_image() -> str: """Return a random image.""" return "" _ = model.bind_tools([random_image]).invoke(messages) def test_anthropic_inputs(self, model: BaseChatModel) -> None: """Test that model can process Anthropic-style message histories. These message histories will include ``AIMessage`` objects with ``tool_use`` content blocks, e.g., .. code-block:: python AIMessage( [ {"type": "text", "text": "Hmm let me think about that"}, { "type": "tool_use", "input": {"fav_color": "green"}, "id": "foo", "name": "color_picker", }, ] ) as well as ``HumanMessage`` objects containing ``tool_result`` content blocks: .. code-block:: python HumanMessage( [ { "type": "tool_result", "tool_use_id": "foo", "content": [ { "type": "text", "text": "green is a great pick! that's my sister's favorite color", # noqa: E501 } ], "is_error": False, }, {"type": "text", "text": "what's my sister's favorite color"}, ] ) This test should be skipped if the model does not support messages of this form (or doesn't support tool calling generally). See Configuration below. .. dropdown:: Configuration To disable this test, set ``supports_anthropic_inputs`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def supports_anthropic_inputs(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that: 1. The model can correctly handle message histories that include message objects with list content. 2. The ``tool_calls`` attribute on AIMessage objects is correctly handled and passed to the model in an appropriate format. 3. HumanMessages with "tool_result" content blocks are correctly handled. Otherwise, if Anthropic tool call and result formats are not supported, set the ``supports_anthropic_inputs`` property to False. """ # noqa: E501 if not self.supports_anthropic_inputs: pytest.skip("Model does not explicitly support Anthropic inputs.") class color_picker(BaseModelV1): """Input your fav color and get a random fact about it.""" fav_color: str human_content: List[dict] = [ { "type": "text", "text": "what's your favorite color in this image", }, ] if self.supports_image_inputs: image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" image_data = base64.b64encode(httpx.get(image_url).content).decode("utf-8") human_content.append( { "type": "image", "source": { "type": "base64", "media_type": "image/jpeg", "data": image_data, }, } ) messages = [ SystemMessage("you're a good assistant"), HumanMessage(human_content), # type: ignore[arg-type] AIMessage( [ {"type": "text", "text": "Hmm let me think about that"}, { "type": "tool_use", "input": {"fav_color": "green"}, "id": "foo", "name": "color_picker", }, ], tool_calls=[ { "name": "color_picker", "args": {"fav_color": "green"}, "id": "foo", "type": "tool_call", } ], ), ToolMessage("That's a great pick!", tool_call_id="foo"), ] response = model.bind_tools([color_picker]).invoke(messages) assert isinstance(response, AIMessage) # Test thinking blocks messages = [ HumanMessage( [ { "type": "text", "text": "Hello", }, ] ), AIMessage( [ { "type": "thinking", "thinking": "I'm thinking...", "signature": "abc123", }, { "type": "text", "text": "Hello, how are you?", }, ] ), HumanMessage( [ { "type": "text", "text": "Well, thanks.", }, ] ), ] response = model.invoke(messages) assert isinstance(response, AIMessage) def test_tool_message_error_status( self, model: BaseChatModel, my_adder_tool: BaseTool ) -> None: """Test that ToolMessage with ``status="error"`` can be handled. These messages may take the form: .. code-block:: python ToolMessage( "Error: Missing required argument 'b'.", name="my_adder_tool", tool_call_id="abc123", status="error", ) If possible, the ``status`` field should be parsed and passed appropriately to the model. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that the ``status`` field on ``ToolMessage`` objects is either ignored or passed to the model appropriately. """ if not self.has_tool_calling: pytest.skip("Test requires tool calling.") model_with_tools = model.bind_tools([my_adder_tool]) messages = [ HumanMessage("What is 1 + 2"), AIMessage( "", tool_calls=[ { "name": "my_adder_tool", "args": {"a": 1}, "id": "abc123", "type": "tool_call", }, ], ), ToolMessage( "Error: Missing required argument 'b'.", name="my_adder_tool", tool_call_id="abc123", status="error", ), ] result = model_with_tools.invoke(messages) assert isinstance(result, AIMessage) def test_message_with_name(self, model: BaseChatModel) -> None: """Test that HumanMessage with values for the ``name`` field can be handled. These messages may take the form: .. code-block:: python HumanMessage("hello", name="example_user") If possible, the ``name`` field should be parsed and passed appropriately to the model. Otherwise, it should be ignored. .. dropdown:: Troubleshooting If this test fails, check that the ``name`` field on ``HumanMessage`` objects is either ignored or passed to the model appropriately. """ result = model.invoke([HumanMessage("hello", name="example_user")]) assert result is not None assert isinstance(result, AIMessage) assert isinstance(result.text(), str) assert len(result.content) > 0 def test_agent_loop(self, model: BaseChatModel) -> None: """Test that the model supports a simple ReAct agent loop. This test is skipped if the ``has_tool_calling`` property on the test class is set to False. This test is optional and should be skipped if the model does not support tool calling (see Configuration below). .. dropdown:: Configuration To disable tool calling tests, set ``has_tool_calling`` to False in your test class: .. code-block:: python class TestMyChatModelIntegration(ChatModelIntegrationTests): @property def has_tool_calling(self) -> bool: return False .. dropdown:: Troubleshooting If this test fails, check that ``bind_tools`` is implemented to correctly translate LangChain tool objects into the appropriate schema for your chat model. Check also that all required information (e.g., tool calling identifiers) from AIMessage objects is propagated correctly to model payloads. This test may fail if the chat model does not consistently generate tool calls in response to an appropriate query. In these cases you can ``xfail`` the test: .. code-block:: python @pytest.mark.xfail(reason=("Does not support tool_choice.")) def test_agent_loop(self, model: BaseChatModel) -> None: super().test_agent_loop(model) """ if not self.has_tool_calling: pytest.skip("Test requires tool calling.") @tool def get_weather(location: str) -> str: """Call to surf the web.""" return "It's sunny." llm_with_tools = model.bind_tools([get_weather]) input_message = HumanMessage("What is the weather in San Francisco, CA?") tool_call_message = llm_with_tools.invoke([input_message]) assert isinstance(tool_call_message, AIMessage) tool_calls = tool_call_message.tool_calls assert len(tool_calls) == 1 tool_call = tool_calls[0] tool_message = get_weather.invoke(tool_call) assert isinstance(tool_message, ToolMessage) response = llm_with_tools.invoke( [ input_message, tool_call_message, tool_message, ] ) assert isinstance(response, AIMessage) @pytest.mark.benchmark @pytest.mark.vcr def test_stream_time( self, model: BaseChatModel, benchmark: BenchmarkFixture, vcr: Cassette ) -> None: """Test that streaming does not introduce undue overhead. See ``enable_vcr_tests`` dropdown :class:`above ` for more information. .. dropdown:: Configuration This test can be enabled or disabled using the ``enable_vcr_tests`` property. For example, to disable the test, set this property to ``False``: .. code-block:: python @property def enable_vcr_tests(self) -> bool: return False .. important:: VCR will by default record authentication headers and other sensitive information in cassettes. See ``enable_vcr_tests`` dropdown :class:`above ` for how to configure what information is recorded in cassettes. """ if not self.enable_vcr_tests: pytest.skip("VCR not set up.") def _run() -> None: for _ in model.stream("Write a story about a cat."): pass if not vcr.responses: _run() else: benchmark(_run) def invoke_with_audio_input(self, *, stream: bool = False) -> AIMessage: """:private:""" raise NotImplementedError() def invoke_with_audio_output(self, *, stream: bool = False) -> AIMessage: """:private:""" raise NotImplementedError() def invoke_with_reasoning_output(self, *, stream: bool = False) -> AIMessage: """:private:""" raise NotImplementedError() def invoke_with_cache_read_input(self, *, stream: bool = False) -> AIMessage: """:private:""" raise NotImplementedError() def invoke_with_cache_creation_input(self, *, stream: bool = False) -> AIMessage: """:private:""" raise NotImplementedError()