Deserialization Error When Loading a Serialized Agent That Includes a Tool

[samirsalman]

Describe the bug
Attempting to load a serialized pipeline that includes an Agent with at least one tool results in a deserialization error. This prevents the use of such agents after serialization, effectively breaking the .run() functionality for pipelines that include tools.

Error message

raise DeserializationError(f"The final attribute is not callable: {attr_value}")
haystack.core.errors.DeserializationError: The final attribute is not callable: Tool(name='get_weather', description='Get the weather for a given location.', parameters={'properties': {'location': {'type': 'string', 'description': 'The location to get the weather for.'}}, 'required': ['location'], 'type': 'object'}, function=<function get_weather at 0x7f23923abd90>, outputs_to_string=None, inputs_from_state=None, outputs_to_state=None)

Expected behavior
The pipeline should be deserialized without errors, and the agent should be fully functional and callable via the .run() method, including the ability to use the tool.

To Reproduce
Run the following Python code to reproduce the issue:

from typing import Annotated

from haystack import Pipeline
from haystack.components.agents import Agent
from haystack.components.generators.chat import OpenAIChatGenerator
from haystack.dataclasses import ChatMessage
from haystack.tools import tool


@tool
def get_weather(location: Annotated[str, "The location to get the weather for."]) -> str:
    "Get the weather for a given location."
    return f"The weather in {location} is sunny with a high of 30°C."


pipeline = Pipeline()

weather_agent = Agent(
    chat_generator=OpenAIChatGenerator(),
    tools=[get_weather],
    system_prompt="""
You are a helpful assistant that can provide weather information.
You can use the get_weather tool to get the weather for a specific location.
""",
)

pipeline.add_component(
    "weather_agent",
    weather_agent,
)

with open("test_serialization.yaml", "w") as f:
    f.write(pipeline.dumps())

pipeline = Pipeline.loads(open("test_serialization.yaml").read())

pipeline.run(
    {
        "weather_agent": {
            "messages": [
                ChatMessage.from_user("What is the weather like in New York?"),
            ]
        }
    }
)

System:

• Haystack version (commit or version number): 2.15.1

[anakin87]

Hello and thanks for reporting this!

I think that @tool converts the function into a Tool with the same name (get_weather), so when you try to reload your Pipeline, the function is no longer available.

Doing the following would work:

from haystack.tools import create_tool_from_function

def get_weather(location: Annotated[str, "The location to get the weather for."]) -> str:
    "Get the weather for a given location."
    return f"The weather in {location} is sunny with a high of 30°C."

weather_tool = create_tool_from_function(get_weather) # use this in your Pipeline

Let me think if we can do something to improve the User Experience of the decorator...

[samirsalman]

Hello @anakin87 , thanks a lot for the suggestion, it works smoothly!

Regarding the @tool decorator: it seems that even when specifying a different name for the tool, deserialization still fails. For example:

@tool(name="_get_weather")
def get_weather(location: Annotated[str, "The location to get the weather for."]) -> str:
    "Get the weather for a given location."
    return f"The weather in {location} is sunny with a high of 30°C."

This appears to be a deserialization issue. In any case, it might be helpful to add a note about this limitation in the documentation, particularly in the section covering the @tool decorator.

Thanks again for your help!