Why:

Enhance OllamaChatGenerator by allowing it to accept a Toolset instance in addition to a list of Tool objects. This brings it in line with other chat generators and improves consistency and flexibility in tool usage.

• fixes Add Toolset support in OllamaChatGenerator #1725

What:

• Updated OllamaChatGenerator to support tools parameter as either a List[Tool] or a Toolset
• Adjusted serialization logic to handle Toolset using serialize_tools_or_toolset
• Modified deserialization to support both formats using deserialize_tools_or_toolset_inplace

How can it be used:

The OllamaChatGenerator can now be used with a Toolset for better tool management and compatibility across generators:

toolset = Toolset(tools)
generator = OllamaChatGenerator(model="qwen3:4b", tools=toolset)

It also allows seamless deserialization and execution with Toolset as part of the component configuration.

How did you test it:

• Added tests for:

  • Initialization with Toolset
  • Serialization and deserialization of Toolset-based components
  • run() execution flow using a Toolset and mocked response from Ollama client
  • Verified backward compatibility with plain list of tools

Manually:

• I've tested qwen3:4b with itinerary agent and bunch of MCPToolset and it kinda worked. Well in some basic form expected from such a relatively small model running on my local CPU.

Notes for the reviewer:

Please confirm:

• Serialization/deserialization logic is sound and consistent with other generators
• ToolCall.id fallback generation doesn't break traceability or downstream logic
• Toolset support is fully covered by new test cases without affecting existing behavior

Note:

• to produce the above referenced Langfuse trace I needed to have add id to ToolCall in OllamaChatGenerator (see 7272bfe diff).
• We'll resolve that issue separately ToolCall ID requirement in to_openai_dict_format causes failure with non-OpenAI models and Langfuse haystack#9418