feat: LlamaStackChatGenerator update tools param to ToolsType

[vblagoje]

Why:

Adopts Haystack's ToolsType to enable flexible tool composition, developers can now mix individual Tool objects and Toolset instances in a single list

part of:

• Update Chat Generators to support mixed Tools and Toolsets #2409

What:

• Updated tools parameter from Union[List[Tool], Toolset] to ToolsType in constructor
• Replaced manual tool handling with Haystack's built-in ToolsType support
• Haystack dependency to 2.19.0 for ToolsType support
• Added tests for mixed tool collection initialization and request parameter formatting

How can it be used:

from haystack.tools import Tool, Toolset
from haystack_integrations.components.generators.llama_stack import LlamaStackChatGenerator

weather_tool = Tool(name="weather", ...)
population_toolset = Toolset([...])

# NEW: Mix Tool and Toolset freely
generator = LlamaStackChatGenerator(tools=[weather_tool, population_toolset])

How did you test it:

• Unit tests verify mixed tool initialization and correct parameter handling
• Backward compatibility validated for existing patterns (list of tools, single toolset)

Notes for the reviewer:

• Centralizes tool normalization logic with Haystack's ToolsType—eliminates duplicate handling code
• Fully backward compatible; no migration needed
• Added integration test test_live_run_with_mixed_tools() demonstrates end-to-end LLM tool invocation from mixed collections

[vblagoje]

Ok after a lot of small fixes - this one should be ready now @anakin87
I think a lot has changed since the last PR, the build for llama stack is now run. And the model naming is different. Perhaps you'd know more about these 🙏

[anakin87]

Looks good from the ToolsType point of view.

However, I'd appreciate it if @Amnah199 could take a look at other changes, since she is more familiar with this integration.

[anakin87]

@Amnah199 please take a look

[vblagoje]

So what was happening is this look at the date stamp - 2 days ago. And it seems like llama stack doesn't have build command any more. So I switched to run command because that one made sense. Using run made a difference and the server started properly.

Then I faced this issue of names. Tests would fail saying there is no such model. I then switched to full naming schema and then everything started to work. cc @anakin87 @Amnah199

[anakin87]

Why are we changing the way we pass the model name? Is it necessary?

[vblagoje]

It seems to be - it was failing for the "llama3.2:3b". I looked at the CI run, the model was booted and ready. Then I tried this naming schema and it worked

[vblagoje]

Looks good from the ToolsType point of view.

However, I'd appreciate it if @Amnah199 could take a look at other changes, since she is more familiar with this integration.

Yes makes sense. @Amnah199 I spent half the day yesterday to get this integration up and running because the old llama stack build command doesn't work anymore. Then I somehow made it run with lamma stack run but the model wasn't recognized without full name. And then with full name it finally started to work.

[vblagoje]

Ah sorry @anakin87 I clicked by accident thinking it was Amna

[Amnah199]

@vblagoje Thanks for the updates. When we created this integration, llama stack was still under development so they were making rapid breaking changes. Hence, all the trouble you had to go through.

[anakin87]

If passing model="llama3.2:3b" is no longer supported, let's also change docstrings for this integration.
Then we would have to update docs and integration page.

[vblagoje]

Ok I'll update the pydocs now as well. 🙏 @anakin87

[anakin87]

👍

Please also open a PR to update https://haystack.deepset.ai/integrations/llama_stack

For some reason, docs are already correct.

[vblagoje]

👍

Please also open a PR to update https://haystack.deepset.ai/integrations/llama_stack

For some reason, docs are already correct.

That's crazy, premonition of a change! We were correct about their naming scheme even before them

[vblagoje]

We need to verify and merge deepset-ai/haystack-integrations#368 cc @bilgeyucel