feat: Add ComponentTool to Haystack tools

[vblagoje]

Why:

Adds ComponentTool enabling Haystack components to be wrapped and used as LLM-compatible tools thus integrating them into Haystack's new tooling architecture.

What:

• Added ComponentTool class to facilitate the use of Haystack components as callable tools.
• Modified __init__.py to include ComponentTool in the module exports.
• Updated pyproject.toml to include the docstring-parser dependency for documentation extraction.

How can it be used:

This implementation enables the creation of LLM-compatible tools from existing components. Here’s an example of usage:

from haystack import component, Pipeline
from haystack.dataclasses import ChatMessage, ChatRole
from haystack.tools import ComponentTool
from haystack.components.tools import ToolInvoker
from haystack.components.generators.chat import OpenAIChatGenerator

@component
class WeatherComponent:
    @component.output_types(reply=str)
    def run(self, city: str, units: str = "celsius"):
        return {"reply":  f"Weather in {city}: 20°{units}"}

weather_tool = ComponentTool(component=WeatherComponent())

# Create pipeline with OpenAIChatGenerator and ToolInvoker
pipeline = Pipeline()
pipeline.add_component("llm", OpenAIChatGenerator(model="gpt-4o-mini", tools=[weather_tool]))
pipeline.add_component("tool_invoker", ToolInvoker(tools=[weather_tool]))

# Connect components
pipeline.connect("llm.replies", "tool_invoker.messages")

message = ChatMessage.from_user("What's the weather like in S.F.?")

# Run pipeline
result = pipeline.run({"llm": {"messages": [message]}})

With weather_tool assigned to your ChatGenerator and ToolInvoker, you can now invoke the component automatically with given a ChatMessage that will trigger this tool, e.g ChatMessage.from_user("What's the weather like in S.F.?")

How did you test it:

Extensive unit and integration tests were created, validating the functionality of ComponentTool with various components. Tests included invoking tools and ensuring correct parameter handling and responses from LLM integrations.

Notes for the reviewer:

Review the implementation of the ComponentTool for potential edge cases with non-component inputs. The integration tests require an OpenAI API key to run successfully, so consider verifying this before executing the tests.

[julian-risch]

Looks quite good to me. My main request is to make the docstring-parser dependency optional. Let's discuss if @anakin87 disagrees or if we decide to make jsonschema a required dependency too.

[julian-risch]

I suggest using lazyimport for this dependency similar to jsonschema.

[julian-risch]

You mentioned limitations in the call regarding which components are supported. Could you please add them to the release note? If there is a particular Haystack component that is not supported mention it here as an example.

[julian-risch]

gpt-4o-mini here too or is it gpt-4 on purpose?

[anakin87]

Here I am contradicting my comment in the experimental PR (https://github.com/deepset-ai/haystack-experimental/pull/159/files/ce864dd8f4f10c2c42f3ded35a515ef71a995585#r1892494701)
but probably now that we have clarified that components to convert into tools MUST NOT be part of a Pipeline, I think it is worth adding a check.

[anakin87]

thanks!

[anakin87]

Let's add a simple test to cover this.

[anakin87]

Our components don't use Pydantic Models, so I’m unsure why we’re supporting this.
Allowing such flexibility might be counterproductive.

If there's an important perspective or requirement that I'm overlooking, I’d appreciate it if you could elaborate.

[anakin87]

Let's add tests for:

• to_dict/from_dict
• complete YAML serde (here is a good example)

[vblagoje]

Thank you @julian-risch and @anakin87 - great feedback. I haven’t removed support for Pydantic models yet but if both of you are okay with its removal - I’ll proceed. The support is largely effortless thanks to TypeAdapter conversion, but as @anakin87 mentioned it could be counterproductive. I’ll remove it in a separate commit to maintain a reference for its implementation in case our users ask for it or an internal need for it arises in the future.

[anakin87]

Let's also remember to add this module to docs/pydoc/config/tools_api.yml

[coveralls]

Pull Request Test Coverage Report for Build 12694409064

Details

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+0.03%) to 91.122%

---

Totals
Change from base Build 12694180152:	0.03%
Covered Lines:	8765
Relevant Lines:	9619

---

💛 - Coveralls