# SPDX-FileCopyrightText: 2022-present deepset GmbH # # SPDX-License-Identifier: Apache-2.0 import json import logging import os from datetime import datetime from typing import Any from unittest.mock import ANY, AsyncMock, MagicMock, patch import pytest from openai import OpenAIError from openai.types.chat import ( ChatCompletion, ChatCompletionChunk, ChatCompletionMessage, ChatCompletionMessageFunctionToolCall, ParsedChatCompletion, ParsedChatCompletionMessage, ParsedChoice, ParsedFunction, ParsedFunctionToolCall, chat_completion_chunk, ) from openai.types.chat.chat_completion import Choice from openai.types.chat.chat_completion_chunk import ChoiceDelta, ChoiceDeltaToolCall, ChoiceDeltaToolCallFunction from openai.types.chat.chat_completion_message_function_tool_call import Function from openai.types.completion_usage import CompletionTokensDetails, CompletionUsage, PromptTokensDetails from pydantic import BaseModel import haystack.components.generators.chat.openai as openai_chat_module from haystack import component from haystack.components.generators.chat.openai import ( OpenAIChatGenerator, _check_finish_reason, _convert_chat_completion_chunk_to_streaming_chunk, _make_schema_strict, ) from haystack.components.generators.utils import print_streaming_chunk from haystack.dataclasses import ( ChatMessage, ChatRole, FileContent, ImageContent, StreamingChunk, ToolCall, ToolCallDelta, ) from haystack.tools import ComponentTool, Tool from haystack.tools.toolset import Toolset from haystack.utils.auth import Secret class CalendarEvent(BaseModel): event_name: str event_date: str event_location: str @pytest.fixture def calendar_event_model(): return CalendarEvent @pytest.fixture def chat_messages(): return [ ChatMessage.from_system("You are a helpful assistant"), ChatMessage.from_user("What's the capital of France"), ] @pytest.fixture def mock_chat_completion_chunk_with_tools(openai_mock_stream): """ Mock the OpenAI API completion chunk response and reuse it for tests """ with patch("openai.resources.chat.completions.Completions.create") as mock_chat_completion_create: completion = ChatCompletionChunk( id="foo", model="gpt-4", object="chat.completion.chunk", choices=[ chat_completion_chunk.Choice( finish_reason="tool_calls", logprobs=None, index=0, delta=chat_completion_chunk.ChoiceDelta( role="assistant", tool_calls=[ chat_completion_chunk.ChoiceDeltaToolCall( index=0, id="123", type="function", function=chat_completion_chunk.ChoiceDeltaToolCallFunction( name="weather", arguments='{"city": "Paris"}' ), ) ], ), ) ], created=int(datetime.now().timestamp()), ) mock_chat_completion_create.return_value = openai_mock_stream( completion, cast_to=None, response=None, client=None ) yield mock_chat_completion_create def weather_function(city: str) -> dict[str, Any]: weather_info = { "Berlin": {"weather": "mostly sunny", "temperature": 7, "unit": "celsius"}, "Paris": {"weather": "mostly cloudy", "temperature": 8, "unit": "celsius"}, "Rome": {"weather": "sunny", "temperature": 14, "unit": "celsius"}, } return weather_info.get(city, {"weather": "unknown", "temperature": 0, "unit": "celsius"}) # mock chat completions with structured outputs @pytest.fixture def mock_parsed_chat_completion(): with patch("openai.resources.chat.completions.Completions.parse") as mock_chat_completion_parse: completion = ParsedChatCompletion[CalendarEvent]( id="json_foo", model="gpt-5-mini", object="chat.completion", choices=[ ParsedChoice[CalendarEvent]( finish_reason="stop", index=0, message=ParsedChatCompletionMessage[CalendarEvent]( content='{"event_name":"Team Meeting","event_date":"2024-03-15",' '"event_location":"Conference Room A"}', refusal=None, role="assistant", annotations=[], audio=None, function_call=None, tool_calls=None, parsed=CalendarEvent( event_name="Team Meeting", event_date="2024-03-15", event_location="Conference Room A" ), ), ) ], created=1757328264, usage=CompletionUsage(completion_tokens=29, prompt_tokens=86, total_tokens=115), ) mock_chat_completion_parse.return_value = completion yield mock_chat_completion_parse @component class MessageExtractor: @component.output_types(messages=list[str], meta=dict[str, Any]) def run(self, messages: list[ChatMessage], meta: dict[str, Any] | None = None) -> dict[str, Any]: """ Extracts the text content of ChatMessage objects :param messages: List of Haystack ChatMessage objects :param meta: Optional metadata to include in the response. :returns: A dictionary with keys "messages" and "meta". """ if meta is None: meta = {} return {"messages": [m.text for m in messages], "meta": meta} @pytest.fixture def tools(): weather_tool = Tool( name="weather", description="useful to determine the weather in a given location", parameters={"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}, function=weather_function, ) # We add a tool that has a more complex parameter signature message_extractor_tool = ComponentTool( component=MessageExtractor(), name="message_extractor", description="Useful for returning the text content of ChatMessage objects", ) return [weather_tool, message_extractor_tool] class TestOpenAIChatGenerator: def test_supported_models(self): """SUPPORTED_MODELS is a non-empty list of strings.""" models = OpenAIChatGenerator.SUPPORTED_MODELS assert isinstance(models, list) assert len(models) > 0 assert all(isinstance(m, str) for m in models) def test_init_default(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") component = OpenAIChatGenerator() assert component.api_key.resolve_value() == "test-api-key" assert component.model == "gpt-5-mini" assert component.streaming_callback is None assert not component.generation_kwargs assert component.timeout is None assert component.max_retries is None assert component.tools is None assert not component.tools_strict assert component.http_client_kwargs is None assert component.client is None assert component.async_client is None def test_init_fail_with_duplicate_tool_names(self, monkeypatch, tools): monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") duplicate_tools = [tools[0], tools[0]] with pytest.raises(ValueError): OpenAIChatGenerator(tools=duplicate_tools) def test_init_with_parameters(self, monkeypatch): tool = Tool(name="name", description="description", parameters={"x": {"type": "string"}}, function=lambda x: x) monkeypatch.setenv("OPENAI_TIMEOUT", "100") monkeypatch.setenv("OPENAI_MAX_RETRIES", "10") component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), streaming_callback=print_streaming_chunk, api_base_url="test-base-url", generation_kwargs={"max_completion_tokens": 10, "some_test_param": "test-params"}, timeout=40.0, max_retries=1, tools=[tool], tools_strict=True, http_client_kwargs={"proxy": "http://example.com:8080", "verify": False}, ) assert component.api_key.resolve_value() == "test-api-key" assert component.model == "gpt-5-mini" assert component.streaming_callback is print_streaming_chunk assert component.generation_kwargs == {"max_completion_tokens": 10, "some_test_param": "test-params"} assert component.timeout == 40.0 assert component.max_retries == 1 assert component.tools == [tool] assert component.tools_strict assert component.http_client_kwargs == {"proxy": "http://example.com:8080", "verify": False} assert component.client is None assert component.async_client is None def test_init_with_parameters_and_env_vars(self, monkeypatch): monkeypatch.setenv("OPENAI_TIMEOUT", "100") monkeypatch.setenv("OPENAI_MAX_RETRIES", "10") component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), streaming_callback=print_streaming_chunk, api_base_url="test-base-url", generation_kwargs={"max_completion_tokens": 10, "some_test_param": "test-params"}, ) assert component.api_key.resolve_value() == "test-api-key" assert component.model == "gpt-5-mini" assert component.streaming_callback is print_streaming_chunk assert component.generation_kwargs == {"max_completion_tokens": 10, "some_test_param": "test-params"} assert component.timeout is None assert component.max_retries is None assert component.client is None assert component.async_client is None def test_to_dict_default(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") component = OpenAIChatGenerator() data = component.to_dict() assert data == { "type": "haystack.components.generators.chat.openai.OpenAIChatGenerator", "init_parameters": { "api_key": {"env_vars": ["OPENAI_API_KEY"], "strict": True, "type": "env_var"}, "model": "gpt-5-mini", "organization": None, "streaming_callback": None, "api_base_url": None, "generation_kwargs": {}, "tools": None, "tools_strict": False, "max_retries": None, "timeout": None, "http_client_kwargs": None, }, } def test_to_dict_with_parameters(self, monkeypatch, calendar_event_model): tool = Tool(name="name", description="description", parameters={"x": {"type": "string"}}, function=print) monkeypatch.setenv("ENV_VAR", "test-api-key") component = OpenAIChatGenerator( api_key=Secret.from_env_var("ENV_VAR"), streaming_callback=print_streaming_chunk, api_base_url="test-base-url", generation_kwargs={ "max_completion_tokens": 10, "some_test_param": "test-params", "response_format": calendar_event_model, "logprobs": True, }, tools=[tool], tools_strict=True, max_retries=10, timeout=100.0, http_client_kwargs={"proxy": "http://example.com:8080", "verify": False}, ) data = component.to_dict() assert data == { "type": "haystack.components.generators.chat.openai.OpenAIChatGenerator", "init_parameters": { "api_key": {"env_vars": ["ENV_VAR"], "strict": True, "type": "env_var"}, "model": "gpt-5-mini", "organization": None, "api_base_url": "test-base-url", "max_retries": 10, "timeout": 100.0, "streaming_callback": "haystack.components.generators.utils.print_streaming_chunk", "generation_kwargs": { "max_completion_tokens": 10, "some_test_param": "test-params", "logprobs": True, "response_format": { "type": "json_schema", "json_schema": { "name": "CalendarEvent", "strict": True, "schema": { "properties": { "event_name": {"title": "Event Name", "type": "string"}, "event_date": {"title": "Event Date", "type": "string"}, "event_location": {"title": "Event Location", "type": "string"}, }, "required": ["event_name", "event_date", "event_location"], "title": "CalendarEvent", "type": "object", "additionalProperties": False, }, }, }, }, "tools": [ { "type": "haystack.tools.tool.Tool", "data": { "async_function": None, "description": "description", "function": "builtins.print", "inputs_from_state": None, "name": "name", "outputs_to_state": None, "outputs_to_string": None, "parameters": {"x": {"type": "string"}}, }, } ], "tools_strict": True, "http_client_kwargs": {"proxy": "http://example.com:8080", "verify": False}, }, } def test_to_dict_with_response_format_json_object(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") component = OpenAIChatGenerator( api_key=Secret.from_env_var("OPENAI_API_KEY"), generation_kwargs={"response_format": {"type": "json_object"}}, ) data = component.to_dict() assert data == { "type": "haystack.components.generators.chat.openai.OpenAIChatGenerator", "init_parameters": { "api_key": {"env_vars": ["OPENAI_API_KEY"], "strict": True, "type": "env_var"}, "model": "gpt-5-mini", "api_base_url": None, "organization": None, "streaming_callback": None, "generation_kwargs": {"response_format": {"type": "json_object"}}, "tools": None, "tools_strict": False, "max_retries": None, "timeout": None, "http_client_kwargs": None, }, } def test_from_dict(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "fake-api-key") data = { "type": "haystack.components.generators.chat.openai.OpenAIChatGenerator", "init_parameters": { "api_key": {"env_vars": ["OPENAI_API_KEY"], "strict": True, "type": "env_var"}, "model": "gpt-5-mini", "api_base_url": "test-base-url", "streaming_callback": "haystack.components.generators.utils.print_streaming_chunk", "max_retries": 10, "timeout": 100.0, "generation_kwargs": {"max_completion_tokens": 10, "some_test_param": "test-params"}, "tools": [ { "type": "haystack.tools.tool.Tool", "data": { "description": "description", "function": "builtins.print", "name": "name", "parameters": {"x": {"type": "string"}}, }, } ], "tools_strict": True, "http_client_kwargs": {"proxy": "http://example.com:8080", "verify": False}, }, } component = OpenAIChatGenerator.from_dict(data) assert isinstance(component, OpenAIChatGenerator) assert component.model == "gpt-5-mini" assert component.streaming_callback is print_streaming_chunk assert component.api_base_url == "test-base-url" assert component.generation_kwargs == {"max_completion_tokens": 10, "some_test_param": "test-params"} assert component.api_key == Secret.from_env_var("OPENAI_API_KEY") assert component.tools == [ Tool(name="name", description="description", parameters={"x": {"type": "string"}}, function=print) ] assert component.tools_strict assert component.timeout == 100.0 assert component.max_retries == 10 assert component.http_client_kwargs == {"proxy": "http://example.com:8080", "verify": False} def test_from_dict_wo_env_var_does_not_fail(self, monkeypatch): monkeypatch.delenv("OPENAI_API_KEY", raising=False) data = { "type": "haystack.components.generators.chat.openai.OpenAIChatGenerator", "init_parameters": { "api_key": {"env_vars": ["OPENAI_API_KEY"], "strict": True, "type": "env_var"}, "model": "gpt-4", "organization": None, "api_base_url": "test-base-url", "streaming_callback": "haystack.components.generators.utils.print_streaming_chunk", "generation_kwargs": {"max_completion_tokens": 10, "some_test_param": "test-params"}, "tools": None, }, } component = OpenAIChatGenerator.from_dict(data) assert component.client is None assert component.async_client is None def test_run(self, chat_messages, openai_mock_chat_completion): component = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key")) response = component.run(chat_messages) # check that the component returns the correct ChatMessage response assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] def test_run_with_string_input(self, openai_mock_chat_completion): component = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key")) response = component.run("What's the capital of France?") _, kwargs = openai_mock_chat_completion.call_args assert kwargs["messages"] == [{"role": "user", "content": "What's the capital of France?"}] assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert isinstance(response["replies"][0], ChatMessage) def test_run_with_params(self, chat_messages, openai_mock_chat_completion): component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), generation_kwargs={"max_completion_tokens": 10, "temperature": 0.5}, ) response = component.run(chat_messages) # check that the component calls the OpenAI API with the correct parameters _, kwargs = openai_mock_chat_completion.call_args assert kwargs["max_completion_tokens"] == 10 assert kwargs["temperature"] == 0.5 # check that the tools are not passed to the OpenAI API (the generator is initialized without tools) assert "tools" not in kwargs # check that the component returns the correct response assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] def test_run_with_params_streaming(self, chat_messages, openai_mock_chat_completion_chunk): streaming_callback_called = False def streaming_callback(chunk: StreamingChunk) -> None: nonlocal streaming_callback_called streaming_callback_called = True component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), streaming_callback=streaming_callback ) response = component.run(chat_messages) # check we called the streaming callback assert streaming_callback_called # check that the component still returns the correct response assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] assert "Hello" in response["replies"][0].text # see openai_mock_chat_completion_chunk def test_run_with_streaming_callback_in_run_method(self, chat_messages, openai_mock_chat_completion_chunk): streaming_callback_called = False def streaming_callback(chunk: StreamingChunk) -> None: nonlocal streaming_callback_called streaming_callback_called = True component = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key")) response = component.run(chat_messages, streaming_callback=streaming_callback) # check we called the streaming callback assert streaming_callback_called # check that the component still returns the correct response assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] assert "Hello" in response["replies"][0].text # see openai_mock_chat_completion_chunk def test_run_with_response_format(self, chat_messages, mock_parsed_chat_completion): component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), generation_kwargs={"response_format": CalendarEvent} ) response = component.run(chat_messages) assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] assert "Team Meeting" in response["replies"][0].text # see mock_parsed_chat_completion def test_run_with_response_format_in_run_method(self, chat_messages, mock_parsed_chat_completion): component = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key")) response = component.run(chat_messages, generation_kwargs={"response_format": CalendarEvent}) assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] assert "Team Meeting" in response["replies"][0].text # see mock_parsed_chat_completion def test_run_with_wrapped_stream_simulation(self, chat_messages, openai_mock_stream): streaming_callback_called = False def streaming_callback(chunk: StreamingChunk) -> None: nonlocal streaming_callback_called streaming_callback_called = True assert isinstance(chunk, StreamingChunk) chunk = ChatCompletionChunk( id="id", model="gpt-4", object="chat.completion.chunk", choices=[chat_completion_chunk.Choice(index=0, delta=chat_completion_chunk.ChoiceDelta(content="Hello"))], created=int(datetime.now().timestamp()), ) # Here we wrap the OpenAI stream in a MagicMock # This is to simulate the behavior of some tools like Weave (https://github.com/wandb/weave) # which wrap the OpenAI stream in their own stream wrapped_openai_stream = MagicMock() wrapped_openai_stream.__iter__.return_value = iter([chunk]) component = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key")) component.warm_up() with patch.object( component.client.chat.completions, "create", return_value=wrapped_openai_stream ) as mock_create: response = component.run(chat_messages, streaming_callback=streaming_callback) mock_create.assert_called_once() assert streaming_callback_called assert "replies" in response assert "Hello" in response["replies"][0].text def test_check_abnormal_completions(self, caplog): caplog.set_level(logging.INFO) messages = [ ChatMessage.from_assistant( "", meta={"finish_reason": "content_filter" if i % 2 == 0 else "length", "index": i} ) for i, _ in enumerate(range(4)) ] for m in messages: _check_finish_reason(m.meta) # check truncation warning message_template = ( "The completion for index {index} has been truncated before reaching a natural stopping point. " "Increase the max_completion_tokens parameter to allow for longer completions." ) for index in [1, 3]: assert caplog.records[index].message == message_template.format(index=index) # check content filter warning message_template = "The completion for index {index} has been truncated due to the content filter." for index in [0, 2]: assert caplog.records[index].message == message_template.format(index=index) def test_run_with_tools(self, tools): with patch("openai.resources.chat.completions.Completions.create") as mock_chat_completion_create: completion = ChatCompletion( id="foo", model="gpt-4", object="chat.completion", choices=[ Choice( finish_reason="tool_calls", logprobs=None, index=0, message=ChatCompletionMessage( role="assistant", tool_calls=[ ChatCompletionMessageFunctionToolCall( id="123", type="function", function=Function(name="weather", arguments='{"city": "Paris"}'), ) ], ), ) ], created=int(datetime.now().timestamp()), usage=CompletionUsage( completion_tokens=40, prompt_tokens=57, total_tokens=97, completion_tokens_details=CompletionTokensDetails( accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0 ), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0), ), ) mock_chat_completion_create.return_value = completion component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), tools=tools[:1], tools_strict=True ) response = component.run([ChatMessage.from_user("What's the weather like in Paris?")]) # ensure that the tools are passed to the OpenAI API function_spec = {**tools[0].tool_spec} function_spec["strict"] = True function_spec["parameters"]["additionalProperties"] = False assert mock_chat_completion_create.call_args[1]["tools"] == [{"type": "function", "function": function_spec}] assert len(response["replies"]) == 1 message = response["replies"][0] assert not message.texts assert not message.text assert message.tool_calls tool_call = message.tool_call assert isinstance(tool_call, ToolCall) assert tool_call.tool_name == "weather" assert tool_call.arguments == {"city": "Paris"} assert message.meta["finish_reason"] == "tool_calls" assert message.meta["usage"]["completion_tokens"] == 40 def test_run_with_tools_and_response_format(self, tools, mock_parsed_chat_completion): """ Test the run method with tools and response format When tools are used, the function call overrides the schema passed in response_format """ with patch("openai.resources.chat.completions.Completions.parse") as mock_chat_completion_parse: completion = ParsedChatCompletion[CalendarEvent]( id="foo", model="gpt-4", object="chat.completion", choices=[ ParsedChoice[CalendarEvent]( finish_reason="tool_calls", logprobs=None, index=0, message=ParsedChatCompletionMessage[CalendarEvent]( role="assistant", tool_calls=[ ParsedFunctionToolCall( id="123", type="function", function=ParsedFunction(name="weather", arguments='{"city": "Paris"}'), ) ], ), ) ], created=int(datetime.now().timestamp()), usage=CompletionUsage( completion_tokens=40, prompt_tokens=57, total_tokens=97, completion_tokens_details=CompletionTokensDetails( accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0 ), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0), ), ) mock_chat_completion_parse.return_value = completion component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), tools=tools[:1], tools_strict=True ) response_with_format = component.run( [ChatMessage.from_user("What's the weather like in Paris?")], generation_kwargs={"response_format": CalendarEvent}, ) assert len(response_with_format["replies"]) == 1 message_with_format = response_with_format["replies"][0] assert not message_with_format.texts assert not message_with_format.text assert message_with_format.tool_calls tool_call = message_with_format.tool_call assert isinstance(tool_call, ToolCall) assert tool_call.tool_name == "weather" assert tool_call.arguments == {"city": "Paris"} assert message_with_format.meta["finish_reason"] == "tool_calls" assert message_with_format.meta["usage"]["completion_tokens"] == 40 def test_run_with_tools_streaming(self, mock_chat_completion_chunk_with_tools, tools): streaming_callback_called = False def streaming_callback(chunk: StreamingChunk) -> None: nonlocal streaming_callback_called streaming_callback_called = True component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), streaming_callback=streaming_callback ) chat_messages = [ChatMessage.from_user("What's the weather like in Paris?")] response = component.run(chat_messages, tools=tools) # check we called the streaming callback assert streaming_callback_called # check that the component still returns the correct response assert isinstance(response, dict) assert "replies" in response assert isinstance(response["replies"], list) assert len(response["replies"]) == 1 assert [isinstance(reply, ChatMessage) for reply in response["replies"]] message = response["replies"][0] assert message.tool_calls tool_call = message.tool_call assert isinstance(tool_call, ToolCall) assert tool_call.tool_name == "weather" assert tool_call.arguments == {"city": "Paris"} assert message.meta["finish_reason"] == "tool_calls" def test_invalid_tool_call_json(self, tools, caplog): caplog.set_level(logging.WARNING) with patch("openai.resources.chat.completions.Completions.create") as mock_create: mock_create.return_value = ChatCompletion( id="test", model="gpt-5-mini", object="chat.completion", choices=[ Choice( finish_reason="tool_calls", index=0, message=ChatCompletionMessage( role="assistant", tool_calls=[ ChatCompletionMessageFunctionToolCall( id="1", type="function", function=Function(name="weather", arguments='"invalid": "json"'), ) ], ), ) ], created=1234567890, usage=CompletionUsage( completion_tokens=47, prompt_tokens=540, total_tokens=587, completion_tokens_details=CompletionTokensDetails( accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0 ), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0), ), ) component = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key"), tools=tools) response = component.run([ChatMessage.from_user("What's the weather in Paris?")]) assert len(response["replies"]) == 1 message = response["replies"][0] assert len(message.tool_calls) == 0 assert "OpenAI returned a malformed JSON string for tool call arguments" in caplog.text assert message.meta["finish_reason"] == "tool_calls" assert message.meta["usage"]["completion_tokens"] == 47 def test_run_with_response_format_and_streaming_pydantic_model(self, calendar_event_model): chat_messages = [ ChatMessage.from_user("The marketing summit takes place on October12th at the Hilton Hotel downtown.") ] component = OpenAIChatGenerator( api_key=Secret.from_token("test-api-key"), generation_kwargs={"response_format": calendar_event_model}, streaming_callback=print_streaming_chunk, ) with pytest.raises(TypeError): component.run(chat_messages) @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run(self): chat_messages = [ChatMessage.from_user("What's the capital of France")] component = OpenAIChatGenerator(model="gpt-4.1-nano", generation_kwargs={"n": 1}) results = component.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] assert "Paris" in message.text assert "gpt-4.1-nano" in message.meta["model"] assert message.meta["finish_reason"] == "stop" assert message.meta["usage"]["prompt_tokens"] > 0 @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_response_format_pydantic_model(self, calendar_event_model): chat_messages = [ ChatMessage.from_user("The marketing summit takes place on October12th at the Hilton Hotel downtown.") ] component = OpenAIChatGenerator( model="gpt-4.1-nano", generation_kwargs={"response_format": calendar_event_model} ) results = component.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] msg = json.loads(message.text) assert "Marketing Summit" in msg["event_name"] assert isinstance(msg["event_date"], str) assert isinstance(msg["event_location"], str) @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_response_format_json_object(self): chat_messages = [ ChatMessage.from_user( 'Answer in JSON: What\'s the capital of France? Please respond with a JSON object with the key "city". ' 'For example: {"city": "Paris"}' ) ] comp = OpenAIChatGenerator(model="gpt-4.1-nano", generation_kwargs={"response_format": {"type": "json_object"}}) results = comp.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] msg = json.loads(message.text) assert "paris" in msg["city"].lower() assert message.meta["finish_reason"] == "stop" @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_response_format_json_object_streaming(self): streaming_callback_called = False def streaming_callback(chunk: StreamingChunk) -> None: nonlocal streaming_callback_called streaming_callback_called = True chat_messages = [ ChatMessage.from_user( 'Answer in JSON: What\'s the capital of France? Please respond with a JSON object with the key "city". ' 'For example: {"city": "Paris"}' ) ] comp = OpenAIChatGenerator( model="gpt-4.1-nano", generation_kwargs={"response_format": {"type": "json_object"}}, streaming_callback=streaming_callback, ) results = comp.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] msg = json.loads(message.text) assert "paris" in msg["city"].lower() assert message.meta["finish_reason"] == "stop" assert streaming_callback_called is True @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_response_format_json_schema(self): response_schema = { "type": "json_schema", "json_schema": { "name": "CapitalCity", "strict": True, "schema": { "title": "CapitalCity", "type": "object", "properties": { "city": {"title": "City", "type": "string"}, "country": {"title": "Country", "type": "string"}, }, "required": ["city", "country"], "additionalProperties": False, }, }, } chat_messages = [ChatMessage.from_user("What's the capital of France?")] comp = OpenAIChatGenerator(model="gpt-4.1-nano", generation_kwargs={"response_format": response_schema}) results = comp.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] msg = json.loads(message.text) assert "Paris" in msg["city"] assert isinstance(msg["country"], str) assert "France" in msg["country"] assert message.meta["finish_reason"] == "stop" @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_response_format_json_schema_streaming(self): streaming_callback_called = False def streaming_callback(chunk: StreamingChunk) -> None: nonlocal streaming_callback_called streaming_callback_called = True response_schema = { "type": "json_schema", "json_schema": { "name": "CapitalCity", "strict": True, "schema": { "title": "CapitalCity", "type": "object", "properties": { "city": {"title": "City", "type": "string"}, "country": {"title": "Country", "type": "string"}, }, "required": ["city", "country"], "additionalProperties": False, }, }, } chat_messages = [ChatMessage.from_user("What's the capital of France?")] comp = OpenAIChatGenerator( model="gpt-4.1-nano", generation_kwargs={"response_format": response_schema}, streaming_callback=streaming_callback, ) results = comp.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] msg = json.loads(message.text) assert "Paris" in msg["city"] assert isinstance(msg["country"], str) assert "France" in msg["country"] assert message.meta["finish_reason"] == "stop" assert streaming_callback_called is True def test_run_with_wrong_model(self): mock_client = MagicMock() mock_client.chat.completions.create.side_effect = OpenAIError("Invalid model name") generator = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key"), model="something-obviously-wrong") generator.client = mock_client with pytest.raises(OpenAIError): generator.run([ChatMessage.from_user("irrelevant")]) @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_streaming(self): class Callback: def __init__(self): self.responses = "" self.counter = 0 def __call__(self, chunk: StreamingChunk) -> None: self.counter += 1 self.responses += chunk.content if chunk.content else "" callback = Callback() component = OpenAIChatGenerator( model="gpt-4.1-nano", streaming_callback=callback, generation_kwargs={"stream_options": {"include_usage": True}}, ) results = component.run([ChatMessage.from_user("What's the capital of France?")]) # Basic response checks assert "replies" in results assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] assert "Paris" in message.text assert isinstance(message.meta, dict) # Metadata checks metadata = message.meta assert "gpt-4.1-nano" in metadata["model"] assert metadata["finish_reason"] == "stop" # Usage information checks assert isinstance(metadata.get("usage"), dict), "meta.usage not a dict" usage = metadata["usage"] assert "prompt_tokens" in usage and usage["prompt_tokens"] > 0 assert "completion_tokens" in usage and usage["completion_tokens"] > 0 # Detailed token information checks assert isinstance(usage.get("completion_tokens_details"), dict), "usage.completion_tokens_details not a dict" assert isinstance(usage.get("prompt_tokens_details"), dict), "usage.prompt_tokens_details not a dict" # Streaming callback verification assert callback.counter > 1 assert "Paris" in callback.responses @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_tools_streaming(self, tools): chat_messages = [ChatMessage.from_user("What's the weather like in Paris and Berlin?")] component = OpenAIChatGenerator( model="gpt-4.1-nano", tools=tools, streaming_callback=print_streaming_chunk, generation_kwargs={"stream_options": {"include_usage": True}}, ) results = component.run(chat_messages) assert len(results["replies"]) == 1 message = results["replies"][0] assert not message.texts assert not message.text assert message.tool_calls tool_calls = message.tool_calls assert len(tool_calls) == 2 for tool_call in tool_calls: assert isinstance(tool_call, ToolCall) assert tool_call.tool_name == "weather" arguments = [tool_call.arguments for tool_call in tool_calls] # Check that both cities are present (case-insensitive, allowing for variations like "Paris, France") city_values = [arg["city"].lower() for arg in arguments] assert any("berlin" in city for city in city_values) assert any("paris" in city for city in city_values) assert message.meta["finish_reason"] == "tool_calls" def test_openai_chat_generator_with_toolset_initialization(self, tools, monkeypatch): """Test that the OpenAIChatGenerator can be initialized with a Toolset.""" monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") toolset = Toolset(tools) generator = OpenAIChatGenerator(tools=toolset) assert generator.tools == toolset def test_from_dict_with_toolset(self, tools, monkeypatch): """Test that the OpenAIChatGenerator can be deserialized from a dictionary with a Toolset.""" monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") toolset = Toolset(tools) component = OpenAIChatGenerator(tools=toolset) data = component.to_dict() deserialized_component = OpenAIChatGenerator.from_dict(data) assert isinstance(deserialized_component.tools, Toolset) assert len(deserialized_component.tools) == len(tools) assert all(isinstance(tool, Tool) for tool in deserialized_component.tools) @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_toolset(self, tools): chat_messages = [ChatMessage.from_user("What's the weather like in Paris?")] toolset = Toolset(tools) component = OpenAIChatGenerator(model="gpt-4.1-nano", tools=toolset) results = component.run(chat_messages) assert len(results["replies"]) == 1 message = results["replies"][0] assert not message.texts assert not message.text assert message.tool_calls tool_call = message.tool_call assert isinstance(tool_call, ToolCall) assert tool_call.tool_name == "weather" assert tool_call.arguments.keys() == {"city"} assert "Paris" in tool_call.arguments["city"] @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_multimodal(self, test_files_path): image_path = test_files_path / "images" / "apple.jpg" # we resize the image to keep this test fast (around 1s) - increase the size in case of errors image_content = ImageContent.from_file_path(file_path=image_path, size=(100, 100), detail="low") chat_messages = [ChatMessage.from_user(content_parts=["What does this image show? Max 5 words", image_content])] generator = OpenAIChatGenerator(model="gpt-4.1-nano") results = generator.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] assert message.text assert "apple" in message.text.lower() assert message.is_from(ChatRole.ASSISTANT) assert not message.tool_calls assert not message.tool_call_results @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_with_file_content(self, test_files_path): pdf_path = test_files_path / "pdf" / "sample_pdf_3.pdf" file_content = FileContent.from_file_path(file_path=pdf_path) chat_messages = [ ChatMessage.from_user( content_parts=[file_content, "Is this document a paper about LLMs? Respond with 'yes' or 'no' only."] ) ] generator = OpenAIChatGenerator(model="gpt-4.1-nano") results = generator.run(chat_messages) assert len(results["replies"]) == 1 message: ChatMessage = results["replies"][0] assert message.is_from(ChatRole.ASSISTANT) assert message.text assert "no" in message.text.lower() def test_init_with_list_of_toolsets(self, monkeypatch, tools): """Test initialization with a list of Toolsets.""" monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") toolset1 = Toolset([tools[0]]) toolset2 = Toolset([tools[1]]) component = OpenAIChatGenerator(tools=[toolset1, toolset2]) assert component.tools == [toolset1, toolset2] assert isinstance(component.tools, list) assert len(component.tools) == 2 assert all(isinstance(ts, Toolset) for ts in component.tools) def test_serde_with_list_of_toolsets(self, monkeypatch, tools): """Test serialization and deserialization with a list of Toolsets.""" monkeypatch.setenv("OPENAI_API_KEY", "test-api-key") toolset1 = Toolset([tools[0]]) toolset2 = Toolset([tools[1]]) component = OpenAIChatGenerator(tools=[toolset1, toolset2]) data = component.to_dict() # Verify serialization preserves list[Toolset] structure tools_data = data["init_parameters"]["tools"] assert isinstance(tools_data, list) assert len(tools_data) == 2 assert all(isinstance(ts, dict) for ts in tools_data) assert tools_data[0]["type"] == "haystack.tools.toolset.Toolset" assert tools_data[1]["type"] == "haystack.tools.toolset.Toolset" # Deserialize and verify deserialized = OpenAIChatGenerator.from_dict(data) assert isinstance(deserialized.tools, list) assert len(deserialized.tools) == 2 assert all(isinstance(ts, Toolset) for ts in deserialized.tools) @pytest.fixture def chat_completion_chunks(): return [ ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[chat_completion_chunk.Choice(delta=ChoiceDelta(role="assistant"), index=0)], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall( index=0, id="call_zcvlnVaTeJWRjLAFfYxX69z4", function=ChoiceDeltaToolCallFunction(arguments="", name="weather"), type="function", ) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='{"ci')) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='ty": ')) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='"Paris')) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='"}'))] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall( index=1, id="call_C88m67V16CrETq6jbNXjdZI9", function=ChoiceDeltaToolCallFunction(arguments="", name="weather"), type="function", ) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='{"ci')) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='ty": ')) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='"Berli')) ] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[ chat_completion_chunk.Choice( delta=ChoiceDelta( tool_calls=[ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='n"}'))] ), index=0, ) ], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[chat_completion_chunk.Choice(delta=ChoiceDelta(), finish_reason="tool_calls", index=0)], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", ), ChatCompletionChunk( id="chatcmpl-BZdwjFecdcaQfCf7bn319vRp6fY8F", choices=[], created=1747834733, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_54eb4bd693", usage=CompletionUsage( completion_tokens=42, prompt_tokens=282, total_tokens=324, completion_tokens_details=CompletionTokensDetails( accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0 ), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0), ), ), ] @pytest.fixture def chat_completion_chunk_delta_none(): chunk = ChatCompletionChunk( id="chatcmpl-BC1y4wqIhe17R8sv3lgLcWlB4tXCw", choices=[chat_completion_chunk.Choice(delta=ChoiceDelta(), index=0)], created=1742207200, model="gpt-5-mini", object="chat.completion.chunk", ) # pydantic complains if we set delta to None at initialization chunk.choices[0].delta = None return chunk @pytest.fixture def streaming_chunks(): return [ StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": None, "finish_reason": None, "received_at": ANY, "usage": None, }, ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ ChoiceDeltaToolCall( index=0, id="call_zcvlnVaTeJWRjLAFfYxX69z4", function=ChoiceDeltaToolCallFunction(arguments="", name="weather"), type="function", ) ], "finish_reason": None, "received_at": ANY, "usage": None, }, index=0, tool_calls=[ToolCallDelta(tool_name="weather", id="call_zcvlnVaTeJWRjLAFfYxX69z4", index=0)], start=True, ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='{"ci'))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=0, tool_calls=[ToolCallDelta(arguments='{"ci', index=0)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='ty": '))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=0, tool_calls=[ToolCallDelta(arguments='ty": ', index=0)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='"Paris'))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=0, tool_calls=[ToolCallDelta(arguments='"Paris', index=0)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction(arguments='"}'))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=0, tool_calls=[ToolCallDelta(arguments='"}', index=0)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ ChoiceDeltaToolCall( index=1, id="call_C88m67V16CrETq6jbNXjdZI9", function=ChoiceDeltaToolCallFunction(arguments="", name="weather"), type="function", ) ], "finish_reason": None, "received_at": ANY, "usage": None, }, index=1, tool_calls=[ToolCallDelta(tool_name="weather", id="call_C88m67V16CrETq6jbNXjdZI9", index=1)], start=True, ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='{"ci'))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=1, tool_calls=[ToolCallDelta(arguments='{"ci', index=1)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='ty": '))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=1, tool_calls=[ToolCallDelta(arguments='ty": ', index=1)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='"Berli'))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=1, tool_calls=[ToolCallDelta(arguments='"Berli', index=1)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": [ChoiceDeltaToolCall(index=1, function=ChoiceDeltaToolCallFunction(arguments='n"}'))], "finish_reason": None, "received_at": ANY, "usage": None, }, index=1, tool_calls=[ToolCallDelta(arguments='n"}', index=1)], ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "index": 0, "tool_calls": None, "finish_reason": "tool_calls", "received_at": ANY, "usage": None, }, finish_reason="tool_calls", ), StreamingChunk( content="", meta={ "model": "gpt-5-mini", "received_at": ANY, "usage": { "completion_tokens": 42, "prompt_tokens": 282, "total_tokens": 324, "completion_tokens_details": { "accepted_prediction_tokens": 0, "audio_tokens": 0, "reasoning_tokens": 0, "rejected_prediction_tokens": 0, }, "prompt_tokens_details": {"audio_tokens": 0, "cached_tokens": 0}, }, }, ), ] @pytest.fixture def mock_openai_clients(monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "fake") sync_cls = MagicMock(name="OpenAI") async_cls = MagicMock(name="AsyncOpenAI") async_cls.return_value.close = AsyncMock() monkeypatch.setattr(openai_chat_module, "OpenAI", sync_cls) monkeypatch.setattr(openai_chat_module, "AsyncOpenAI", async_cls) return sync_cls, async_cls class TestComponentLifecycle: def test_warm_up_uses_default_timeout_and_max_retries(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "fake-api-key") generator = OpenAIChatGenerator() generator.warm_up() assert generator.client.max_retries == 5 assert generator.client.timeout == 30.0 def test_warm_up_uses_timeout_and_max_retries_from_parameters(self): generator = OpenAIChatGenerator(api_key=Secret.from_token("fake-api-key"), timeout=40.0, max_retries=1) generator.warm_up() assert generator.client.max_retries == 1 assert generator.client.timeout == 40.0 def test_warm_up_uses_timeout_and_max_retries_from_env_vars(self, monkeypatch): monkeypatch.setenv("OPENAI_TIMEOUT", "100") monkeypatch.setenv("OPENAI_MAX_RETRIES", "10") generator = OpenAIChatGenerator(api_key=Secret.from_token("fake-api-key")) generator.warm_up() assert generator.client.max_retries == 10 assert generator.client.timeout == 100.0 def test_key_resolved_at_warm_up_not_init(self, monkeypatch): monkeypatch.delenv("OPENAI_API_KEY", raising=False) generator = OpenAIChatGenerator() with pytest.raises(ValueError, match="None of the .* environment variables are set"): generator.warm_up() def test_warm_up_warms_tools_once(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "fake-api-key") warm_up_calls = [] class MockTool(Tool): def __init__(self, tool_name): super().__init__( name=tool_name, description=f"Mock tool {tool_name}", parameters={"type": "object", "properties": {"x": {"type": "string"}}, "required": ["x"]}, function=lambda x: x, ) def warm_up(self): warm_up_calls.append(self.name) generator = OpenAIChatGenerator(tools=[MockTool("tool1"), MockTool("tool2")]) assert not generator._tools_warmed_up generator.warm_up() assert sorted(warm_up_calls) == ["tool1", "tool2"] assert generator._tools_warmed_up generator.warm_up() assert sorted(warm_up_calls) == ["tool1", "tool2"] def test_warm_up_with_no_tools_does_not_raise(self, monkeypatch): monkeypatch.setenv("OPENAI_API_KEY", "fake-api-key") generator = OpenAIChatGenerator() generator.warm_up() assert generator._tools_warmed_up def test_sync_lifecycle(self, mock_openai_clients): sync_cls, _ = mock_openai_clients generator = OpenAIChatGenerator() assert generator.client is None assert generator.async_client is None generator.warm_up() assert generator.client is sync_cls.return_value assert generator.async_client is None generator.close() sync_cls.return_value.close.assert_called_once() assert generator.client is None async def test_async_lifecycle(self, mock_openai_clients): _, async_cls = mock_openai_clients generator = OpenAIChatGenerator() await generator.warm_up_async() assert generator.async_client is async_cls.return_value assert generator.client is None await generator.close_async() async_cls.return_value.close.assert_awaited_once() assert generator.async_client is None async def test_close_is_safe_without_warm_up(self, mock_openai_clients): generator = OpenAIChatGenerator() generator.close() await generator.close_async() assert generator.client is None assert generator.async_client is None async def test_close_and_close_async_are_independent(self, mock_openai_clients): generator = OpenAIChatGenerator() generator.warm_up() await generator.warm_up_async() generator.close() assert generator.client is None assert generator.async_client is not None await generator.close_async() assert generator.async_client is None class TestChatCompletionChunkConversion: def test_convert_chat_completion_chunk_to_streaming_chunk(self, chat_completion_chunks, streaming_chunks): previous_chunks = [] for openai_chunk, haystack_chunk in zip(chat_completion_chunks, streaming_chunks, strict=True): stream_chunk = _convert_chat_completion_chunk_to_streaming_chunk( chunk=openai_chunk, previous_chunks=previous_chunks ) assert stream_chunk == haystack_chunk previous_chunks.append(stream_chunk) def test_convert_chat_completion_chunk_with_empty_tool_calls(self): # This can happen with some LLM providers where tool calls are not present but the pydantic models are still # initialized. chunk = ChatCompletionChunk( id="chatcmpl-BC1y4wqIhe17R8sv3lgLcWlB4tXCw", choices=[ chat_completion_chunk.Choice( delta=chat_completion_chunk.ChoiceDelta( tool_calls=[ChoiceDeltaToolCall(index=0, function=ChoiceDeltaToolCallFunction())] ), index=0, ) ], created=1742207200, model="gpt-5-mini", object="chat.completion.chunk", ) result = _convert_chat_completion_chunk_to_streaming_chunk(chunk=chunk, previous_chunks=[]) assert result.content == "" assert result.start is False assert result.tool_calls == [ToolCallDelta(index=0)] assert result.tool_call_result is None assert result.index == 0 assert result.meta["model"] == "gpt-5-mini" assert result.meta["received_at"] is not None def test_convert_chat_completion_chunk_with_delta_none(self, chat_completion_chunk_delta_none): """ Test that a chat completion chunk with a delta set to None is converted to a streaming chunk properly. This should not happen, but some OpenAI-compatible providers sometimes return a delta set to None. """ result = _convert_chat_completion_chunk_to_streaming_chunk( chunk=chat_completion_chunk_delta_none, previous_chunks=[] ) assert result.content == "" assert result.start is False assert result.tool_calls is None assert result.tool_call_result is None assert result.index == 0 assert result.component_info is None assert result.finish_reason is None assert result.reasoning is None assert result.meta["model"] == "gpt-5-mini" assert result.meta["received_at"] is not None assert result.meta["index"] == 0 assert result.meta["finish_reason"] is None assert result.meta["usage"] is None assert result.meta["tool_calls"] is None def test_handle_stream_response(self, chat_completion_chunks, chat_completion_chunk_delta_none): openai_chunks = [chat_completion_chunk_delta_none] + chat_completion_chunks comp = OpenAIChatGenerator(api_key=Secret.from_token("test-api-key")) result = comp._handle_stream_response(openai_chunks, callback=lambda _: None)[0] # type: ignore assert not result.texts assert not result.text # Verify both tool calls were found and processed assert len(result.tool_calls) == 2 assert result.tool_calls[0].id == "call_zcvlnVaTeJWRjLAFfYxX69z4" assert result.tool_calls[0].tool_name == "weather" assert result.tool_calls[0].arguments == {"city": "Paris"} assert result.tool_calls[1].id == "call_C88m67V16CrETq6jbNXjdZI9" assert result.tool_calls[1].tool_name == "weather" assert result.tool_calls[1].arguments == {"city": "Berlin"} # Verify meta information assert result.meta["model"] == "gpt-5-mini" assert result.meta["finish_reason"] == "tool_calls" assert result.meta["index"] == 0 assert result.meta["completion_start_time"] is not None assert result.meta["usage"] == { "completion_tokens": 42, "prompt_tokens": 282, "total_tokens": 324, "completion_tokens_details": { "accepted_prediction_tokens": 0, "audio_tokens": 0, "reasoning_tokens": 0, "rejected_prediction_tokens": 0, }, "prompt_tokens_details": {"audio_tokens": 0, "cached_tokens": 0}, } def test_convert_usage_chunk_to_streaming_chunk(self): usage_chunk = ChatCompletionChunk( id="chatcmpl-BC1y4wqIhe17R8sv3lgLcWlB4tXCw", choices=[], created=1742207200, model="gpt-5-mini", object="chat.completion.chunk", service_tier="default", system_fingerprint="fp_06737a9306", usage=CompletionUsage( completion_tokens=8, prompt_tokens=13, total_tokens=21, completion_tokens_details=CompletionTokensDetails( accepted_prediction_tokens=0, audio_tokens=0, reasoning_tokens=0, rejected_prediction_tokens=0 ), prompt_tokens_details=PromptTokensDetails(audio_tokens=0, cached_tokens=0), ), ) result = _convert_chat_completion_chunk_to_streaming_chunk(chunk=usage_chunk, previous_chunks=[]) assert result.content == "" assert result.start is False assert result.tool_calls is None assert result.tool_call_result is None assert result.meta["model"] == "gpt-5-mini" assert result.meta["received_at"] is not None class TestMakeSchemaStrict: def test_flat_object(self): schema = {"type": "object", "properties": {"name": {"type": "string"}}} result = _make_schema_strict(schema) assert result == { "type": "object", "properties": {"name": {"type": "string"}}, "additionalProperties": False, "required": ["name"], } def test_nested_object(self): schema = { "type": "object", "properties": { "person": {"type": "object", "properties": {"name": {"type": "string"}, "age": {"type": "integer"}}} }, } result = _make_schema_strict(schema) assert result == { "type": "object", "properties": { "person": { "type": "object", "properties": {"name": {"type": "string"}, "age": {"type": "integer"}}, "additionalProperties": False, "required": ["name", "age"], } }, "additionalProperties": False, "required": ["person"], } def test_defs_and_ref(self): schema = { "type": "object", "properties": {"address": {"$ref": "#/$defs/Address"}}, "$defs": { "Address": {"type": "object", "properties": {"street": {"type": "string"}, "city": {"type": "string"}}} }, } result = _make_schema_strict(schema) assert result == { "type": "object", "properties": {"address": {"$ref": "#/$defs/Address"}}, "$defs": { "Address": { "type": "object", "properties": {"street": {"type": "string"}, "city": {"type": "string"}}, "additionalProperties": False, "required": ["street", "city"], } }, "additionalProperties": False, "required": ["address"], } def test_array_items(self): schema = { "type": "object", "properties": { "people": {"type": "array", "items": {"type": "object", "properties": {"name": {"type": "string"}}}} }, } result = _make_schema_strict(schema) assert result == { "type": "object", "properties": { "people": { "type": "array", "items": { "type": "object", "properties": {"name": {"type": "string"}}, "additionalProperties": False, "required": ["name"], }, } }, "additionalProperties": False, "required": ["people"], } def test_anyof(self): schema = { "type": "object", "properties": { "value": {"anyOf": [{"type": "string"}, {"type": "object", "properties": {"x": {"type": "integer"}}}]} }, } result = _make_schema_strict(schema) assert result == { "type": "object", "properties": { "value": { "anyOf": [ {"type": "string"}, { "type": "object", "properties": {"x": {"type": "integer"}}, "additionalProperties": False, "required": ["x"], }, ] } }, "additionalProperties": False, "required": ["value"], } def test_does_not_mutate_original(self): schema = {"type": "object", "properties": {"a": {"type": "string"}}} result = _make_schema_strict(schema) assert "additionalProperties" not in schema assert "required" not in schema assert result == { "type": "object", "properties": {"a": {"type": "string"}}, "additionalProperties": False, "required": ["a"], } def test_preserves_existing_required(self): schema = { "type": "object", "properties": {"a": {"type": "string"}, "b": {"type": "integer"}}, "required": ["a"], } result = _make_schema_strict(schema) assert result == { "type": "object", "properties": {"a": {"type": "string"}, "b": {"type": "integer"}}, "additionalProperties": False, "required": ["a", "b"], } def test_complex_schema_with_defs_and_combinators(self): schema = { "type": "object", "properties": { "messages": {"type": "array", "items": {"$ref": "#/$defs/ChatMessage"}}, "config": { "oneOf": [ {"type": "null"}, { "type": "object", "properties": {"temperature": {"type": "number"}, "max_tokens": {"type": "integer"}}, }, ] }, }, "$defs": { "ChatMessage": { "type": "object", "properties": { "role": {"type": "string"}, "content": {"anyOf": [{"type": "string"}, {"type": "null"}]}, "meta": { "type": "object", "properties": { "model": {"type": "string"}, "usage": { "type": "object", "properties": { "prompt_tokens": {"type": "integer"}, "completion_tokens": {"type": "integer"}, }, }, }, }, }, } }, } result = _make_schema_strict(schema) assert result == { "type": "object", "properties": { "messages": {"type": "array", "items": {"$ref": "#/$defs/ChatMessage"}}, "config": { "oneOf": [ {"type": "null"}, { "type": "object", "properties": {"temperature": {"type": "number"}, "max_tokens": {"type": "integer"}}, "additionalProperties": False, "required": ["temperature", "max_tokens"], }, ] }, }, "$defs": { "ChatMessage": { "type": "object", "properties": { "role": {"type": "string"}, "content": {"anyOf": [{"type": "string"}, {"type": "null"}]}, "meta": { "type": "object", "properties": { "model": {"type": "string"}, "usage": { "type": "object", "properties": { "prompt_tokens": {"type": "integer"}, "completion_tokens": {"type": "integer"}, }, "additionalProperties": False, "required": ["prompt_tokens", "completion_tokens"], }, }, "additionalProperties": False, "required": ["model", "usage"], }, }, "additionalProperties": False, "required": ["role", "content", "meta"], } }, "additionalProperties": False, "required": ["messages", "config"], } def test_prepare_api_call_strict_nested_tool(self): nested_tool = Tool( name="create_person", description="Create a person record", parameters={ "type": "object", "properties": { "name": {"type": "string"}, "address": { "type": "object", "properties": {"street": {"type": "string"}, "city": {"type": "string"}}, }, }, "required": ["name"], }, function=lambda name, address: f"{name} at {address}", ) component = OpenAIChatGenerator(api_key=Secret.from_token("test-key"), tools_strict=True) api_args = component._prepare_api_call(messages=[ChatMessage.from_user("test")], tools=[nested_tool]) tool_def = api_args["tools"][0]["function"] assert tool_def["strict"] is True assert tool_def["parameters"] == { "type": "object", "properties": { "name": {"type": "string"}, "address": { "type": "object", "properties": {"street": {"type": "string"}, "city": {"type": "string"}}, "additionalProperties": False, "required": ["street", "city"], }, }, "additionalProperties": False, "required": ["name", "address"], } @pytest.mark.skipif( not os.environ.get("OPENAI_API_KEY", None), reason="Export an env var called OPENAI_API_KEY containing the OpenAI API key to run this test.", ) @pytest.mark.integration def test_live_run_strict_nested_tool(self): tool = Tool( name="create_person", description="Create a person record with an address", parameters={ "type": "object", "properties": { "name": {"type": "string", "description": "Full name"}, "address": { "type": "object", "properties": { "street": {"type": "string", "description": "Street address"}, "city": {"type": "string", "description": "City name"}, }, }, }, }, function=lambda name, address: f"{name} at {address}", ) component = OpenAIChatGenerator(model="gpt-4.1-nano", tools_strict=True) results = component.run( messages=[ChatMessage.from_user("Create a person named John at 123 Main St, Springfield")], tools=[tool] ) assert len(results["replies"]) == 1 message = results["replies"][0] assert message.tool_calls tool_call = message.tool_call assert tool_call.tool_name == "create_person" assert "name" in tool_call.arguments assert "address" in tool_call.arguments assert "street" in tool_call.arguments["address"] assert "city" in tool_call.arguments["address"]