feat: HuggingFaceLocalChatGenerator unified support for tools

[vblagoje]

Why:

Adds universal tools support to HuggingFaceLocalChatGenerator

• fixes HuggingFaceLocal ChatGenerator - support for Tool #8731

What:

• Introduced support for tool calls, allowing the generation model to prepare and parse tool interactions.
• Implemented a tool call parsing method that utilizes regex patterns or callables.

How can it be used:

The addition allows for parsing tool calls from generated text, identified by tool names and their arguments. This can be set up during initialization:

import torch

from haystack.dataclasses import ChatMessage
from haystack.tools import Tool
from haystack.components.generators.chat import HuggingFaceLocalChatGenerator
from haystack.utils import Secret

def weather(city: str):
    return f'The weather in {city} is sunny and 32°C'

# Define tool parameters
tool_parameters = {
    "type": "object",
    "properties": {"city": {"type": "string"}},
    "required": ["city"]
}

# Create weather tool
weather_tool = Tool(
    name="weather",
    description="useful to determine the weather in a given location",
    parameters=tool_parameters,
    function=weather
)
# Initialize generator with tool
client = HuggingFaceLocalChatGenerator(
    model="NousResearch/Hermes-2-Pro-Llama-3-8B",
    token=Secret.from_env_var("HF_TOKEN"),
    tools=[weather_tool],
    huggingface_pipeline_kwargs={"torch_dtype":torch.bfloat16, "device_map": "auto"}
)
client.warm_up()

# Run initial query
messages = [ChatMessage.from_user("What's the weather like in Paris?")]
results = client.run(messages=messages)

# Get tool call from response
tool_message = next(msg for msg in results["replies"] if msg.tool_call)
tool_call = tool_message.tool_call

# Execute tool and send result back
weather_result = weather(**tool_call.arguments)
new_messages = [
    messages[0],
    tool_message,
    ChatMessage.from_tool(tool_result=weather_result, origin=tool_call)
]

# Get final response
final_result = client.run(new_messages)
print(final_result["replies"][0].text)

>>> The weather in Paris is sunny and 32°C. Enjoy your day!

How did you test it:

• Added unit tests that mock generator run with specified tool patterns and tool lists
• Real test notebook

Notes for the reviewer:

• See https://huggingface.co/blog/unified-tool-use for more details about HF and tools
• Validate the regex pattern for tool call extraction.
• Ensure comprehensive testing of different tool call scenarios to guarantee robustness.

[coveralls]

Pull Request Test Coverage Report for Build 13236543453

Warning: This coverage report may be inaccurate.

This pull request's base commit is no longer the HEAD commit of its target branch. This means it includes changes from outside the original pull request, including, potentially, unrelated coverage changes.

• For more information on this, see Tracking coverage changes with pull request builds.
• To avoid this issue with future PRs, see these Recommended CI Configurations.
• For a quick fix, rebase this PR at GitHub. Your next report should be accurate.

Details

• 0 of 0 changed or added relevant lines in 0 files are covered.
• 24 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-0.4%) to 92.228%

---

Files with Coverage Reduction	New Missed Lines	%
telemetry/_telemetry.py	1	84.15%
components/generators/chat/hugging_face_local.py	23	84.46%

Totals
Change from base Build 13183889468:	-0.4%
Covered Lines:	9173
Relevant Lines:	9946

---

💛 - Coveralls

[vblagoje]

@davidsbatista asking @anakin87 for review here as he knows this code intimately 🙏

[anakin87]

Nice work! I particularly appreciated the notebook...

I left some comments about points we can improve.

[vblagoje]

Have a look at the updated run of the notebook as well @anakin87 🙏

[anakin87]

I left a small comment to address.

Plus, "feat!:" in the PR title indicates a breaking change, and this does not seem breaking.
If I am right, let's simply change this to "feat:"

Feel free to merge once these small aspects are fixed.

Good work!

[anakin87]

this generated_text is confusing. Please use something like "mock response. We don't use it"

[anakin87]

we are only extracting a single tool call, but this is clearly explained.
We can improve this in future if there are requests from the community.

[vblagoje]

Yes exactly, that's why I left it as a list 💪